Publications

Accurate phenological stage classification is crucial for addressing global challenges to food security posed by climate change, water scarcity, and land degradation. It enables precision agriculture by optimizing key interventions such as irrigation, fertilization, and pest control. While deep learning offers powerful tools, existing methods face four key limitations: reliance on narrow features and models, limited long-term forecasting capability, computational inefficiency, and opaque, unvalidated explanations. To overcome these limitations, this paper presents a deep learning framework for phenology classification, utilizing multi-source time series data from satellite imagery, meteorological stations, and field observations. The approach emphasizes temporal consistency, spatial adaptability, computational efficiency, and explainability. A feature engineering pipeline extracts temporal dynamics via lag features, rolling statistics, Fourier transforms and seasonal encodings. Feature selection combines incremental strategies with classical filter, wrapper, and embedded methods. Deep learning models across multiple paradigms-feedforward, recurrent, convolutional, and attention-based-are benchmarked under multi-horizon forecasting tasks. To reduce model complexity while preserving performance where possible, the framework employs knowledge distillation, transferring predictive knowledge from complex teacher models to compact and deployable student models. For model interpretability, a new Hybrid SHAP-Association Rule Explainability approach is proposed, integrating model-driven and data-driven explanations. Agreement between views is quantified using trust metrics: precision@k, coverage, and Jaccard similarity, with a retraining-based validation mechanism. Experiments on phenology data from Andalusia demonstrate high accuracy, strong generalizability, trustworthy explanations and resource-efficient phenology monitoring in agricultural systems.

In the literature, several studies have shown that state-of-the-art image similarity metrics are not perceptual metrics; moreover, they have difficulty evaluating images, especially when texture distortion is also present. In this work, we propose a new perceptual metric composed of two terms. The first term evaluates the dissimilarity between the textures of two images using Earth Mover's Distance. The second term evaluates the chromatic dissimilarity between two images in the Oklab perceptual color space. We evaluated the performance of our metric on a non-traditional dataset, called Berkeley-Adobe Perceptual Patch Similarity, which contains a wide range of complex distortions in shapes and colors. We have shown that our metric outperforms the state of the art, especially when images contain shape distortions, confirming also its greater perceptiveness. Furthermore, although deep black-box metrics could be very accurate, they only provide similarity scores between two images, without explaining their main differences and similarities. Our metric, on the other hand, provides visual explanations to support the calculated score, making the similarity assessment transparent and justified.

The eXplainable Artificial Intelligence (XAI) research predominantly concentrates on providing explanations about AI model decisions, especially Deep Learning (DL) models. However, there is a growing interest in using XAI techniques to automatically improve the performance of the AI systems themselves. This paper proposes IMPACTX, a novel approach that leverages XAI as a fully automated attention mechanism, without requiring external knowledge or human feedback. Experimental results show that IMPACTX improves predictive performance compared to standalone baseline models by integrating XAI-based supervision provided by teacher explanations into the training process. Furthermore, IMPACTX directly provides new feature attribution maps for the model's decisions, without relying on external XAI methods during the inference process. Our proposal is evaluated using three widely recognized DL models (EfficientNet-B2, MobileNet, and LeNet-5) and on six publicly available benchmark datasets, comprising three image datasets (CIFAR-10, CIFAR-100, and STL-10) and three tabular datasets (Covertype, CDC Diabetes Health Indicators, and Adult). The results show that IMPACTX consistently improves the performance of all the inspected DL models across all evaluated datasets, and it directly provides appropriate explanations for its responses.

This paper concerns methodology for exploiting the multi-objective Extremal Optimization for load-balancing algorithms in high-performance distributed systems. In clusters and data centers, there has always been a trade-off between contradictory goals such as obtaining high performance, reducing inter-node communication, task or virtual machine migration, and energy savings. Thus, a multi-objective optimization strategy should be provided based on task migration to achieve an efficient processor load balance in the executive distributed environment, which is an NP-hard computational problem. The paper proposes a new selection scheme for the final load-balanced solution in the Pareto front. In this gradient-supported scheme, we examine lexicographic solutions relaxed by a margin of allowable loss, provided that the remaining optimization criteria are improved. This has been achieved by calculating the gradients of the tangent lines connecting the analyzed lexicographic solutions and the subsequent Pareto front points. The algorithm has been evaluated by comparative simulation experiments with application program graphs run in distributed systems. The evaluation, which included a comparison with a genetic algorithm, confirmed the very good performance of the proposed gradient-based Pareto front selection method.

Artificial intelligence and deep learning are powerful tools for extracting knowledge from large particularly in healthcare. However, their black-box nature raises interpretability concerns, especially stakes applications. Existing eXplainable Artificial Intelligence methods often focus solely on visualization rule-based explanations, limiting interpretability's depth and clarity. This work proposes a novel AI method specifically designed for medical image analysis, integrating statistical, visual, and explanations to improve transparency in deep learning models. Statistical features are derived features extracted using a custom Mobilenetv2 model. A two-step feature selection method-filtering with mutual importance selection-ranks and refines these features. Decision tree and RuleFit are employed to classify data and extract human-readable rules. Additionally, a novel statistical feature overlay visualization generates heatmap-like representations of three key statistical measures (mean, and entropy), providing both localized and quantifiable visual explanations of model decisions. The method has been validated on five medical imaging datasets-COVID-19 radiography, ultrasound breast brain tumor magnetic resonance imaging, lung and colon cancer histopathological, and glaucoma images results confirmed by medical experts, demonstrating its effectiveness in enhancing interpretability image classification tasks.

Breast cancer is the most common cancer in women overall and the fifth leading cause of cancer mortality worldwide (1). Artificial Intelligence (AI) plays a prominent role, providing the necessary tools for the development of predictive models crucial to support the decision-making process of the involved clinicians. Early detection and characterization of cancer are of utmost importance in reducing breast cancer mortality and improving clinical outcome and life quality. In this scenario, the ability to have AI-based predictive models that can provide clues as to the location of the tumor will allow to address the disease when it is still in its early stages of development and, potentially, allow clinicians to make decisive and more conservative interventions for the patient.

Early and accurate detection of lemon disease is necessary for effective citrus crop management. Traditional approaches often lack refined diagnosis, necessitating more powerful solutions. The article introduces adaptive PSO-LemonNetX, a novel framework integrating a novel deep learning model, adaptive Particle Swarm Optimization (PSO)-based feature selection, and explainable AI (XAI) using LIME. The approach improves the accuracy of classification while also enhancing the explainability of the model. Our end-to-end model obtained 97.01% testing and 98.55% validation accuracy. Performance was enhanced further with adaptive PSO and conventional classifiers-100% validation accuracy using Naive Bayes and 98.8% testing accuracy using Naive Bayes and an SVM. The suggested PSO-based feature selection performed better than ReliefF, Kruskal-Wallis, and Chi-squared approaches. Due to its lightweight design and good performance, this approach can be adapted for edge devices in IoT-enabled smart farms, contributing to sustainable and automated disease detection systems. These results show the potential of integrating deep learning, PSO, grid search, and XAI into smart agriculture workflows for enhancing agricultural disease detection and decision-making.

Circulating tumor cells (CTCs) are cells that shed from primary tumors into the bloodstream, leading to the formation of metastases. Their presence and numbers reflect the tumor burden, making them valuable biomarkers. Consequently, a simple blood draw can be used to non-invasively monitor tumor progression and treatment efficacy, a rapidly growing approach known as liquid biopsy [1]. However, the extreme rarity and heterogeneity of CTCs pose significant challenges for their identification, requiring methods that are often time-consuming, complex, and expensive. Building on the concept that cytomorphological criteria can differentiate between healthy and unhealthy cells, we have developed an innovative microscope on a chip designed for automated, high-throughput imaging of samples flowing in a microfluidic channel. Utilizing advanced laser-based microfabrication techniques, namely femtosecond laser micromachining [2], we have integrated optical splitter, and a 3D optical remapper into a glass substrate assembled together with fiber-based optical delay lines (as in the scheme of Fig.1a). This system operates by splitting a single nanosecond laser pulse into a sequence of pulses that illuminate the specimen at different locations and times (Fig.1b illustrates device characterization). The signals are then efficiently collected using a fast photodetector and a high speed digitizer for subsequent intensity variation analysis, enabling the acquisition of a single cell image in just a few microseconds. This optical system is intended for detecting specimens flowing within a microfluidic channel. The fluidic layout has been optimized by employing 3D hydrodynamic focusing, in which sheath flows confine the sample within a narrow region precisely aligned with the detection area [3]. This optofluidic device has been successfully validated using calibration beads and various types of tumor cells, demonstrating its reliability and versatility (Fig.1c). This high level of integration ensures user-friendly operation and provides consistent image quality across different days and conditions, a fundamental aspects when processing rare cells as CTCs. The captured images are fully compatible with machine learning algorithms, enabling efficient and accurate classification of cells. This approach has the potential to transform the detection and analysis of CTCs, providing a scalable, cost-effective solution for advancing liquid biopsy technologies.

The increasing use of black-box networks in high-risk contexts has led researchers to propose explainable methods to make these networks transparent. Most methods that allow us to understand the behavior of Deep Neural Networks (DNNs) are post-hoc approaches, implying that the explainability is questionable, as these methods do not clarify the internal behavior of a model. Thus, this demonstrates the difficulty of interpreting the internal behavior of deep models. This systematic literature review collects the ante-hoc methods that provide an understanding of the internal mechanisms of deep models and which can be helpful to researchers who need to use interpretability methods to clarify DNNs. This work provides the definitions of strong interpretability and weak interpretability, which will be used to describe the interpretability of the methods discussed in this article. The results of this work are divided mainly into prototype-based methods, concept-based methods, and other interpretability methods for deep models.

Circulating Tumor Cells (CTCs) are crucial biomarkers in liquid biopsy, offering a noninvasive tool for cancer patient management. However, their identification remains particularly challenging due to their limited number and heterogeneity. Labeling samples for contrast limits the generalization of fluorescence-based methods across different hospital datasets. Analyzing single-cell images enables detailed assessment of cell morphology, subcellular structures, and phenotypic variations, often hidden in clustered images. Developing a method based on bright-field single-cell analysis could overcome these limitations. CTCs can be isolated using an unbiased workflow combining Parsortix® technology, which selects cells based on size and deformability, with DEPArray™ technology, enabling precise visualization and selection of single cells. Traditionally, DEPArray-acquired digital images are manually analyzed, making the process time-consuming and prone to variability. In this study, we present a Deep Learning-based (DL) classification pipeline designed to distinguish CTCs from leukocytes in blood samples, aimed to enhance diagnostic accuracy and optimize clinical workflows. Our approach employs images from the bright-field channel acquired through DEPArray technology leveraging a ResNet-based Convolutional Neural Network. To improve model generalization, we applied three types of data augmentation techniques and incorporated fluorescence (DAPI) channel images into the training phase, allowing the network to learn additional CTC-specific features. Notably, only bright-field images have been used for testing, ensuring the model’s ability to identify CTCs without relying on fluorescence markers. The proposed model achieved an F1-score of 0.798, demonstrating its capability to distinguish CTCs from leukocytes. These findings highlight the potential of DL in refining CTC analysis and advancing liquid biopsy applications.

Breast cancer classification through ultrasound imaging poses a significant challenge due to the inherent noise present in ultrasound images. The radiologist’s reporting process aims to assess the lesions within the images following the Breast Imaging-Reporting and Data System (BI-RADS). This work investigates whether the medical knowledge, represented by the BI-RADS information, augmented by pixel-based quantitative features, can improve breast cancer classification. Machine learning classifiers, including XGBoost, Random Forest, and Support Vector Machine, were trained with an intelligible multimodal signature composed of the BI-RADS and radiomic features. Exploiting the intrinsic interpretability of our model input, the work aims to obtain an explainable predictive model using post-hoc explanation methods. A proprietary dataset composed of 237 B-mode ultrasound scans was acquired, and a total of 103 radiomics features were extracted. Before the training of classifiers, a pipeline for selecting an informative and non-redundant signature was implemented. A 10-fold Cross-Validation repeated 20 times was considered for the training in 80% of the dataset, and the best model in terms of accuracy was selected for testing on the remaining 20%. Our results prove how the medical knowledge, represented by the BI-RADS information, is enhanced with the use of radiomic features. XGBoost was the best model, showing an AUROC of 0.977 .± 0.029 and 0.956 in the training and test phases, respectively. In addition, the implemented global explanation using the SHAP method and exploiting the intelligibility of radiomic features, allowed us to confirm some important model findings.

The fragmentation of patient data across different and disconnected systems, ranging from electronic health records to Artificial Intelligence (AI)-based diagnostic tools, poses a major challenge to the delivery of an efficient and accurate healthcare system. This paper proposes a modular and interoperable archi- tecture designed to integrate heterogeneous clinical data from different sources, including structured clinical records, socio- health information, patient-generated data, and outputs from AI- based diagnostic systems such as imaging analysis. The proposed architecture facilitates seamless data harmo- nization and supports clinical decision-making by structuring integrated information through the HL7 Fast Healthcare Interop- erability Resources (FHIR) standard. This enables standardized data exchange and full interoperability with existing Health Information Systems, including Electronic Health Records and Telemedicine Platforms. An Implementation Guide is proposed as a reference framework for validating the FHIR resources produced by the architecture. In addition, a key feature of the architecture is its embedded Clinical Decision Support System, which dynamically identifies and presents only the clinically relevant information required for diagnostic reasoning and risk assessment.

In the past few years, Federated Learning (FL) has emerged as an effective approach for training Neural Networks (NNs) over a computing network while preserving data privacy. Most existing FL approaches require defining a priori 1) a predefined structure for all the NNs running on the clients and 2) an explicit aggregation procedure. These can be limiting factors in cases where pre-defining such algorithmic details is difficult. Recently, NEvoFed was proposed, an FL method that leverages Neuroevolution running on the clients, in which the NN structures are heterogeneous and the aggregation is implicitly accomplished on the client side. Here, we propose MFC-NEvoFed, a novel approach to FL that does not require learning models, i.e., neural network parameters, to be distributed over the networks, thus taking a step towards security improvement. The only information exchanged in client/server communication is the performance of each model on local data, allowing the emergence of optimal NN architectures without needing any kind of model aggregation. Another appealing feature of our framework is that it can be used with any Machine Learning algorithm provided that, during the learning phase, the model updates do not depend on the input data. To assess the validity of MFC-NEvoFed, we test it on four datasets, showing that very compact NNs can be obtained without drops in performance compared to canonical FL. Finally, such compact structures allow for a step towards explainability, which is highly desirable in domains such as digital health, from which the tested datasets come.

The widespread fragmentation of patient information across heterogeneous systems and the lack of standardized integration mechanisms hinder efficient and comprehensive medi- cal diagnostics. To address these limitations, this work presents an architecture framework designed to support physicians in the diagnostic process by integrating clinical and socio- health information (patient medical histories), structured documents extracted from Health Information System (HIS), and data automatically extracted from diagnostic images using Artificial Intelligence (AI) techniques. The proposed architecture is made by several mod- ules, in particular a Decision Support System (DSS) that enables risk assessment related to specific patient’s clinical conditions. In addition, the clinical information retrieved is aggregated, standardized, and transmitted to external systems for follow up. Standard- ization and data interoperability are ensured through the adoption of the international HL7 Fast Healthcare Interoperability Resources (FHIR) standard, which facilitates seamless connection with HIS. An Android application has been developed to communicate with different HISs in order to: i) retrieve information, ii) aggregate clinical data, iii) calculate patient risk scores using AI algorithms, iv) display results to healthcare professionals, and v) generate and share relevant clinical information with external systems in a standardized format. To demonstrate architecture’s applicability, a case study on breast cancer diagnosis is presented. In this context, an AI-based Risk Assessment module was developed using the Breast Ultrasound Images Dataset (BUSI), which includes benign, malignant, and nor- mal cases. Machine Learning algorithms were applied to perform the classification task. Model performance was evaluated using a 4-fold cross-validation strategy to ensure robust- ness and generalizability. The best results were achieved using the Multilayer Perceptron method, with a competitive F1-score of 0.97.

This work presents an integrated, high-throughput microscope on a chip designed for rapid and automated circulating tumor cells imaging based on cytomorphological features. The system employs a modified time-stretch imag- ing technique, utilizing a single nanosecond laser pulse split into a sequence of temporally and spatially separated pulses to illuminate the whole cell at different moments. Fabricated using femtosecond laser micromachining, the device inte- grates optical circuits, delay lines, and a microfluidic chip, enabling high-speed image acquisition with a single-pixel detector. The system is validated using calibration beads and tumor cells, demonstrating high resolution and stability. Fully compatible with machine learning algorithms, this platform represents a scalable, cost-effective solution for advancing real-time liquid biopsy and can- cer diagnostics.

In recent years, architectural and algorithmic innovations in machine learning have revolutionized the analysis of medical images [...]

Multimodal datasets offer valuable support for developing Clinical Decision Support Systems (CDSS), which leverage predictive models to enhance clinicians' decision-making. In this observational study, we present a dataset of suspected Coronary Artery Disease (CAD) patients - called MultiD4CAD - comprising imaging and clinical data. The imaging data obtained from Coronary Computed Tomography Angiography (CCTA) includes epicardial (EAT) and pericoronary (PAT) adipose tissue segmentations. These metabolically active fat tissues play a key role in cardiovascular diseases. In addition, clinical data include a set of biomarkers recognized as CAD risk factors. The validated EAT and PAT segmentations make the dataset suitable for training predictive models based on radiomics and deep learning architectures. The inclusion of CAD disease labels allows for its application in supervised learning algorithms to predict CAD outcomes. MultiD4CAD has revealed important correlations between imaging features, clinical biomarkers, and CAD status. The article concludes by discussing some challenges, such as classification, segmentation, radiomics, and deep training tasks, that can be investigated and validated using the proposed dataset.

To support pathologists in breast tumor diagnosis, deep learning plays a crucial role in the development of histological whole slide image (WSI) classification methods. However, automatic classification is challenging due to the high-resolution data and the scarcity of representative training data. To tackle these limitations, we propose a deep learning-based breast tumor gigapixel histological image multi-classifier integrated with a high-resolution data augmentation model to process the entire slide by exploring its local and global information and generating its different synthetic versions. The key idea is to perform the classification and augmentation in feature latent space, reducing the computational cost while preserving the class label of the input. We adopt a deep learning-based multi-classification method and evaluate the contribution given by a conditional generative adversarial network-based data augmentation model on the classifier’s performance for three tumor classes in the BRIGHT Challenge dataset. The proposed method has allowed us to achieve an average F1 equal to 69.5, considering only the WSI dataset of the Challenge. The results are comparable to those obtained by the Challenge winning method (71.6), also trained on the annotated tumor region dataset of the Challenge.

This research presents DeepCRINet, a deep learning (DL) model designed for reliable performance across various Chest Radiography Images (CRIs) datasets, in response to the urgent need for quick and accurate lung disease identification utilizing CRIs. Our method builds on earlier research, which frequently used single-source datasets that might not adequately represent the heterogeneity present in clinical situations. Our model's diagnostic adaptability and real-world dependability are improved by utilizing images from different datasets, which helps us overcome limitations such as dataset bias, robustness, generalizability, and underrepresentation of conditions. With validation on a broad dataset consisting of 14,096 images (from three different datasets), DeepCRINet provides a solution that demonstrates excellent flexibility in recognizing illnesses including TuBerculosis, Pneumonia, COVID-19, and Lung Opacity. Through data augmentation, we improve the dataset, supporting training and testing procedures and confirming the model's ability to generalize. We used occlusion sensitivity as a kind of explainable AI to openly identify and visually emphasize regions important to proper classification. This ability not only shows that DeepCRINet is analytically better than other DL models and hybrid techniques, but it also improves patient outcomes and diagnosis, which makes it a vital tool for medical professionals like radiologists.

The use of Super-Resolution SR algorithms applied to Magnetic Resonance Images (MRIs) is increasingly common in the medical field. Increasing the resolution of images allows physicians to more easily observe image details. Over the years, several SR approaches have been tried by researchers. Among the various approaches, Diffusion Models (DMs) have been shown to perform well in the SR task. In this work, we propose the use of a Latent Diffusion Model (LDM) for the SR of medical images. Different studies have shown that LDMs improve the performance of DMs in several SR tasks. To our knowledge, LDMs have not been tested for SR of medical images such as MRIs. We therefore perform fine-tuning of an LDM on medical datasets. To evaluate the SR images generated by the LDM, we compare them to the original high-resolution images using two similarity measurements. We show that the LDM achieves better similarity values than other SR models on the same medical dataset. We also show with visual examples the advantage of applying SR using an LDM.

There is an urgent need in many application areas for eXplainable ArtiiciaI Intelligence (XAI) approaches to boost people’s conidence and trust in Artiicial Intelligence methods. Current works concentrate on speciic aspects of XAI and avoid a comprehensive perspective. This study undertakes a systematic survey of importance, approaches, methods, and application domains to address this gap and provide a comprehensive understanding of the XAI domain. Applying the Systematic Literature Review approach has resulted in inding and discussing 155 papers, allowing a wide discussion on the strengths, limitations, and challenges of XAI methods and future research directions.

The problem of executing machine learning algorithms over data while complying with data privacy is highly relevant in many application areas, including medicine in general and Precision Medicine in particular. In this paper, an innovative framework for Federated Learning is proposed that allows performing machine learning and effectively tackling the issue of data privacy while taking a step towards security during communication. Unlike the standard federated approaches where models should travel on the communication networks and would be subject to possible cyberattacks, the models proposed by our framework do not need to travel, thus moving in the direction of security improvement. Another very appealing feature is that it can be used with any machine learning algorithm provided that, during the learning phase, the model updating does not depend on the input data. To show its effectiveness, the learning process is here accomplished by an Evolutionary Algorithm, namely Grammatical Evolution, thus also obtaining explicit knowledge that can be provided to the domain experts to justify the decisions made. As a test case, glucose values prediction for a number of patients with type 1 diabetes is considered and is tackled as a classification problem, the goal being to predict for any future value a possible range. Finally, a comparison of the performance of the proposed framework is performed against that of a non-Federated Learning approach.

In the past few years, Federated Learning (FL) has emerged as an effective approach for training neural networks (NNs) over a computing network while preserving data privacy. Most of the existing FL approaches require the user to define a priori the same structure for all the NNs running on the clients, along with an explicit aggregation procedure. This can be a limiting factor in cases where pre-defining such algorithmic details is difficult. To overcome these issues, we propose a novel approach to FL, which leverages Neuroevolution running on the clients. This implies that the NN structures may be different across clients, hence providing better adaptation to the local data. Furthermore, in our approach, the aggregation is implicitly accomplished on the client side by exploiting the information about the models used on the other clients, thus allowing the emergence of optimal NN architectures without needing an explicit aggregation. We test our approach on three datasets, showing that very compact NNs can be obtained without significant drops in performance compared to canonical FL. Moreover, we show that such compact structures allow for a step towards explainability, which is highly desirable in domains such as digital health, from which the tested datasets come.

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disease affecting motorneurons of the bulbar, cervical, thoracic, or lumbar segments. Bulbar presentation is a devastating characteristic that impairs patients' ability to communicate and is linked to shorter survival. Early acoustic manifestation of voice symptoms, such as dysarthria, is very variable, making its detection and classification challenging, both by human specialists and automatic systems. In this context, precision medicine, defined as "prevention and treatment strategies that take individual variability into account", has gained a great interest in the ALS community. Specifically, the use of innovative Artificial Intelligence techniques, such as Machine Learning, plays a pivotal role in finding specific patterns in the data set to help neurologists in clinical decision-making. Therefore, the main objective of this study was to find new markers, and new patterns, to promptly detect the possible presence of dysarthria and to correctly classify its severity. We have performed an acoustic analysis on different voice signals of various degrees of impairment acquired during outpatient visits at the ALS center of the "Federico II" University Hospital. From the collected signals, a new database containing different acoustic parameters was realized, on which several experiments were performed. The study led us to the discovery of markers that helped to develop a decision tree that separated healthy subjects from patients and, among patients, those with different severity of dysarthria. This model achieved good results in terms of dysarthria classification accuracy, 86.6%, which is excellent considering the small number of subjects in the data set.

This paper describes a new publicly-available database of VOiCe signals acquired in Amyotrophic Lateral Sclerosis (ALS) patients (VOC-ALS) and healthy controls performing different speech tasks. This dataset consists of 1224 voice signals recorded from 153 participants: 51 healthy controls (32 males and 19 females) and 102 ALS patients (65 males and 37 females) with different severity of dysarthria. Each subject’s voice was recorded using a smartphone application (Vox4Health) while performing several vocal tasks, including a sustained phonation of the vowels /a/, /e/, /i/, /o/, /u/ and /pa/, /ta/, /ka/ syllable repetition. Basic derived speech metrics such as harmonics-to-noise ratio, mean and standard deviation of fundamental frequency (F0), jitter and shimmer were calculated. The F0 standard deviation of vowels and syllables showed an excellent ability to identify people with ALS and to discriminate the different severity of dysarthria. These data represent the most comprehensive database of voice signals in ALS and form a solid basis for research on the recognition of voice impairment in ALS patients for use in clinical applications.

This study examines the use of different interpretable Artificial Intelligence models in predicting short-term blood glucose levels in subjects with Type 1 Diabetes. The interpretability of Artificial Intelligence models is a critical concept, especially in the medical context, because it prevents the development of the so-called “black boxes” and provides decisions that are fully understandable by both patients and healthcare professionals. The final aim of this work is to integrate such fully comprehensible models within a glucose monitoring system to ensure a more transparent management of insulin therapy and an improved patient adherence. The predictive ability of the models has been assessed using a dataset containing glucose levels and heart rate variability features for certain patients selected from the open D1NAMO dataset. The prediction problem was initially approached as a multi-series regression issue and then re-evaluated as a problem of accurate classification into seven glycemic ranges. Evaluating the predictive abilities of the models in terms of correct classifications, we show that Decision Tree outperforms the other models for the analyzed subjects, achieving a weighted F1 score of 0.87 for the best run. Finally, the experiments have also shown that integrating heart rate variability features opens up the possibility of developing non-invasive monitoring systems, reducing the burden on patients and improving their quality of life.

Purpose: To investigate the feasibility of an artificial intelligence (AI)-based semi-automated segmentation for the extraction of ultrasound (US)-derived radiomics features in the characterization of focal breast lesions (FBLs). Material and methods: Two expert radiologists classified according to US BI-RADS criteria 352 FBLs detected in 352 patients (237 at Center A and 115 at Center B). An AI-based semi-automated segmentation was used to build a machine learning (ML) model on the basis of B-mode US of 237 images (center A) and then validated on an external cohort of B-mode US images of 115 patients (Center B). Results: A total of 202 of 352 (57.4%) FBLs were benign, and 150 of 352 (42.6%) were malignant. The AI-based semi-automated segmentation achieved a success rate of 95.7% for one reviewer and 96% for the other, without significant difference (p = 0.839). A total of 15 (4.3%) and 14 (4%) of 352 semi-automated segmentations were not accepted due to posterior acoustic shadowing at B-Mode US and 13 and 10 of them corresponded to malignant lesions, respectively. In the validation cohort, the characterization made by the expert radiologist yielded values of sensitivity, specificity, PPV and NPV of 0.933, 0.9, 0.857, 0.955, respectively. The ML model obtained values of sensitivity, specificity, PPV and NPV of 0.544, 0.6, 0.416, 0.628, respectively. The combined assessment of radiologists and ML model yielded values of sensitivity, specificity, PPV and NPV of 0.756, 0.928, 0.872, 0.855, respectively. Conclusion: AI-based semi-automated segmentation is feasible, allowing an instantaneous and reproducible extraction of US-derived radiomics features of FBLs. The combination of radiomics and US BI-RADS classification led to a potential decrease of unnecessary biopsy but at the expense of a not negligible increase of potentially missed cancers.

Feature extraction and selection from medical data are the basis of radiomics and image biomarker discovery for various architectures, including convolutional neural networks (CNNs). We herein describe the typical radiomics steps and the components of a CNN for both deep feature extraction and end-to-end approaches. We discuss the curse of dimensionality, along with dimensionality reduction techniques. Despite the outstanding performance of deep learning (DL) approaches, the use of handcrafted features instead of deep learned features needs to be considered for each specific study. Dataset size is a key factor: large-scale datasets with low sample diversity could lead to overfitting; limited sample sizes can provide unstable models. The dataset must be representative of all the “facets” of the clinical phenomenon/disease investigated. The access to the high-performance computational resources from graphics processing units is another key factor, especially for the training phase of deep architectures. The advantages of multi-institutional federated/collaborative learning are described. When large language models are used, high stability is needed to avoid catastrophic forgetting in complex domain-specific tasks. We highlight that non-DL approaches provide model explainability superior to that provided by DL approaches. To implement explainability, the need for explainable AI arises, also through post hoc mechanisms.

This paper presents an innovative approach leveraging Neuronal Manifold Analysis of EEG data to identify specific time intervals for feature extraction, effectively capturing both class-specific and subject-specific characteristics. Different pipelines were constructed and employed to extract distinctive features within these intervals, specifically for motor imagery (MI) tasks. The methodology was validated using the Graz Competition IV datasets 2A (four-class) and 2B (two-class) motor imagery classification, demonstrating an improvement in classification accuracy that surpasses state-of-the-art algorithms designed for MI tasks. A multi-dimensional feature space, constructed using NMA, was built to detect intervals that capture these critical characteristics, which led to significantly enhanced classification accuracy, especially for individuals with initially poor classification performance. These findings highlight the robustness of this method and its potential to improve classification performance in EEG-based MI-BCI systems.

Breast Cancer (BC) is the most common cancer among women worldwide and is characterized by intra- and inter-tumor heterogeneity that strongly contributes towards its poor prognosis. The Estrogen Receptor (ER), Progesterone Receptor (PR), Human Epidermal Growth Factor Receptor 2 (HER2), and Ki67 antigen are the most examined markers depicting BC heterogeneity and have been shown to have a strong impact on BC prognosis. Radiomics can noninvasively predict BC heterogeneity through the quantitative evaluation of medical images, such as Magnetic Resonance Imaging (MRI), which has become increasingly important in the detection and characterization of BC. However, the lack of comprehensive BC datasets in terms of molecular outcomes and MRI modalities, and the absence of a general methodology to build and compare feature selection approaches and predictive models, limit the routine use of radiomics in the BC clinical practice. In this work, a new radiomic approach based on a two-step feature selection process was proposed to build predictors for ER, PR, HER2, and Ki67 markers. An in-house dataset was used, containing 92 multiparametric MRIs of patients with histologically proven BC and all four relevant biomarkers available. Thousands of radiomic features were extracted from post-contrast and subtracted Dynamic Contrast-Enanched (DCE) MRI images, Apparent Diffusion Coefficient (ADC) maps, and T2-weighted (T2) images. The two-step feature selection approach was used to identify significant radiomic features properly and then to build the final prediction models. They showed remarkable results in terms of F1-score for all the biomarkers: 84%, 63%, 90%, and 72% for ER, HER2, Ki67, and PR, respectively. When possible, the models were validated on the TCGA/TCIA Breast Cancer dataset, returning promising results (F1-score = 88% for the ER+/ER− classification task). The developed approach efficiently characterized BC heterogeneity according to the examined molecular biomarkers.

Deep Learning based heart sound classification is of significant interest in reducing the burden of manual auscultation through the automated detection of signals, including abnormal heartbeats. This work presents a method for classifying phonocardiogram (PCG) signals as normal or abnormal by applying a deep Convolutional Neural Network (CNN) after transforming the signals into 2D color images. In particular, a new methodology based on fractal theory, which exploits Partitioned Iterated Function Systems (PIFS) to generate 2D color images from 1D signals is presented. PIFS have been extensively investigated in the context of image coding and indexing on account of their ability to interpolate and identify self-similar features in an image. Our classification approach has shown a high potential in terms of noise robustness and does not require any pre-processing steps or an initial segmentation of the signal, as instead happens in most of the approaches proposed in the literature. In this preliminary work, we have carried out several experiments on the database released for the 2016 Physionet Challenge, both in terms of different classification networks and different inputs to the networks, thus also evaluating the data quality. Among all experiments, we have obtained the best result of 0.85 in terms of modified Accuracy (MAcc).

Colon and Lung cancers are two of the most common causes of mortality in adults. They may simultaneously form in organs and have a detrimental effect on human life. There is a high risk that cancer will spread to the two organs if it is not discovered in the early stages. One of the most essential elements of successful therapy is the histological diagnosis of such cancers. Deep learning algorithms have improved the speed and accuracy of time-consuming and challenging procedures, enabling researchers to examine a huge number of patients swiftly and inexpensively. By examining their histological images and applying modern deep learning, this study develops a classification framework called DeepLCCNet to discriminate between five kinds of colon and lung tissues (three malignant and two benign). More precisely, we have classified five tissue types of Lung and Colon Cancer Histopathological Images data set using our model, i.e., benign tissue of the lung, squamous cell carcinoma of the lung, adenocarcinoma of the lung, benign tissue of the colon, and adenocarcinoma of the colon. According to the results, the proposed model can detect cancer tissues with an average accuracy of 99.67% and maximum accuracy of 99.84%. Medical professionals will be able to utilize a precise, automated system for detecting and classifying various kinds of colon and lung cancers.

In this paper, we propose an innovative Federated Learning-inspired evolutionary frame- work. Its main novelty is that this is the first time that an Evolutionary Algorithm is employed on its own to directly perform Federated Learning activity. A further novelty resides in the fact that, differently from the other Federated Learning frameworks in the literature, ours can efficiently deal at the same time with two relevant issues in Machine Learning, i.e., data privacy and interpretability of the solutions. Our framework consists of a master/slave approach in which each slave contains local data, protecting sensible private data, and exploits an evolutionary algorithm to generate prediction models. The master shares through the slaves the locally learned models that emerge on each slave. Sharing these local models results in global models. Being that data privacy and interpretability are very significant in the medical domain, the algorithm is tested to forecast future glucose values for diabetic patients by exploiting a Grammatical Evolution algorithm. The effectiveness of this knowledge-sharing process is assessed experimentally by comparing the proposed framework with another where no exchange of local models occurs. The results show that the performance of the proposed approach is better and demonstrate the validity of its sharing process for the emergence of local models for personal diabetes management, usable as efficient global models. When further subjects not involved in the learning process are considered, the models discovered by our framework show higher generalization capability than those achieved without knowledge sharing: the improvement provided by knowledge sharing is equal to about 3.03% for precision, 1.56% for recall, 3.17% for F1, and 1.56% for accuracy. Moreover, statistical analysis reveals the statistical superiority of model exchange with respect to the case of no exchange taking place.

Diabetes mellitus is a metabolic disease involving high blood glucose levels that can lead to serious medical consequences. Hence, for diabetic patients the prediction of future glucose levels is essential in the management of the disease. Most of the forecasting approaches in the literature evaluate the effectiveness of glucose predictors only with numerical metrics. These approaches are limited because they evenly treat all the errors without considering their different clinical impact that could involve lethal effects in dangerous situations such as hypo- or hyperglycemia. To overcome such a limitation, this paper aims to devise models for reducing high-risk glucose forecasting errors for Type 1 diabetic patients. For this purpose, we exploit a Grammatical Evolution algorithm to induce personalized and interpretable forecasting glucose models assessed with a novel, composite metric to satisfy both clinical and numerical requirements of the estimated predictions. To assess the effectiveness of the proposed approach, a real-world data set widely used in literature, consisting of data from several patients suffering from Type 1 diabetes, has been adopted. The experimental findings show that the induced models are interpretable and capable of assuring predictions with a good tradeoff between medical quality and numerical accuracy and with remarkable performance in reducing high-risk glucose forecasting errors. Furthermore, their performance is better than or comparable to that of other state-of-the-art methods.

Purpose: This study was aimed to investigate the predictive value of the radiomics features extracted from pericoronaric adipose tissue around the anterior interventricular artery (IVA) - to assess the condition of coronary arteries compared with the use of clinical characteristics alone (i.e., risk factors). Methods: Clinical and radiomic data of 118patients were retrospectively analyzed. In total, 93 radiomics features were extracted for each ROI around the IVA, and 13 clinical features were used to build different machine learning models finalized to predict the impairment (or otherwise) of coronary arteries. Results: Pericoronaric radiomic features improved prediction above the use of risk factors alone. In fact, with the best model (Random Forest + Mutual Information) the AUROC reached 0.820 ± 0.076. As matter of fact, the combined use of both types of features (i.e., radiomic and clinical) allows for improved performance regardless of the feature selection method used. Conclusion: Experimental findings demonstrated that the use of radiomic features alone achieves better performance than the use of clinical features alone, while the combined use of both clinical and radiomic biomarkers further improves the predictive ability of the models. The main contribution of this work concerns: i) the implementation of multimodal predictive models, based on both clinical and radiomic features, and ii) a trusted system to support clinical decision-making processes by means of explainable classifiers and interpretable features.

The SARS-CoV-2 virus pandemic had devastating effects on various aspects of life: clinical cases, ranging from mild to severe, can lead to lung failure and to death. Due to the high incidence, data-driven models can support physicians in patient management. The explainability and interpretability of machine-learning models are mandatory in clinical scenarios. In this work, clinical, laboratory and radiomic features were used to train machine-learning models for COVID-19 prognosis prediction. Using Explainable AI algorithms, a multi-level explainable method was proposed taking into account the developer and the involved stakeholder (physician, and patient) perspectives. A total of 1023 radiomic features were extracted from 1589 Chest X-Ray images (CXR), combined with 38 clinical/laboratory features. After the pre-processing and selection phases, 40 CXR radiomic features and 23 clinical/laboratory features were used to train Support Vector Machine and Random Forest classifiers exploring three feature selection strategies. The combination of both radiomic, and clinical/laboratory features enabled higher performance in the resulting models. The intelligibility of the used features allowed us to validate the models' clinical findings. According to the medical literature, LDH, PaO2 and CRP were the most predictive laboratory features. Instead, ZoneEntropy and HighGrayLevelZoneEmphasis - indicative of the heterogeneity/uniformity of lung texture - were the most discriminating radiomic features. Our best predictive model, exploiting the Random Forest classifier and a signature composed of clinical, laboratory and radiomic features, achieved AUC=0.819, accuracy=0.733, specificity=0.705, and sensitivity=0.761 in the test set. The model, including a multi-level explainability, allows us to make strong clinical assumptions, confirmed by the literature insights.

Radiomic analysis allows for the detection of imaging biomarkers supporting decision-making processes in clinical environments, from diagnosis to prognosis. Frequently, the original set of radiomic features is augmented by considering high-level features, such as wavelet transforms. However, several wavelets families (so called kernels) are able to generate different multi-resolution representations of the original image, and which of them produces more salient images is not yet clear. In this study, an in-depth analysis is performed by comparing different wavelet kernels and by evaluating their impact on predictive capabilities of radiomic models. A dataset composed of 1589 chest X-ray images was used for COVID-19 prognosis prediction as a case study. Random forest, support vector machine, and XGBoost were trained (on a subset of 1103 images) after a rigorous feature selection strategy to build-up the predictive models. Next, to evaluate the models generalization capability on unseen data, a test phase was performed (on a subset of 486 images). The experimental findings showed that Bior1.5, Coif1, Haar, and Sym2 kernels guarantee better and similar performance for all three machine learning models considered. Support vector machine and random forest showed comparable performance, and they were better than XGBoost. Additionally, random forest proved to be the most stable model, ensuring an appropriate balance between sensitivity and specificity.

Melanoma is a malignant cancer type which develops when DNA damage occurs (mainly due to environmental factors such as ultraviolet rays). Often, melanoma results in intense and aggressive cell growth that, if not caught in time, can bring one toward death. Thus, early identification at the initial stage is fundamental to stopping the spread of cancer. In this paper, a ViT-based architecture able to classify melanoma versus non-cancerous lesions is presented. The proposed predictive model is trained and tested on public skin cancer data from the ISIC challenge, and the obtained results are highly promising. Different classifier configurations are considered and analyzed in order to find the most discriminating one. The best one reached an accuracy of 0.948, sensitivity of 0.928, specificity of 0.967, and AUROC of 0.948.

Introduction The aim of our study was to evaluate the feasibility of texture analysis of epicardial fat (EF) and thoracic subcutaneous fat (TSF) in patients undergoing cardiac CT (CCT). Materials and methods We compared a consecutive population of 30 patients with BMI <=25 kg/m2 (Group A, 60.6 ± 13.7 years) with a control population of 30 patients with BMI >25 kg/m2 (Group B, 63.3 ± 11 years). A dedicated computer application for quantification of EF and a texture analysis application for the study of EF and TSF were employed. Results The volume of EF was higher in group B (mean 116.1 cm3 vs. 86.3 cm3, p = 0.014), despite no differences were found neither in terms of mean density (-69.5 ± 5 HU vs. -68 ± 5 HU, p = 0.28), nor in terms of quartiles distribution (Q1, p = 0.83; Q2, p = 0.22, Q3, p = 0.83, Q4, p = 0.34). The discriminating parameters of the histogram class were mean (p = 0.02), 0,1st (p = 0.001), 10th (p = 0.002), and 50th percentiles (p = 0.02). DifVarnc was the discriminating parameter of the co-occurrence matrix class (p = 0.007). The TSF thickness was 15 ± 6 mm in group A and 19.5 ± 5 mm in group B (p = 0.003). The TSF had a mean density of -97 ± 19 HU in group A and -95.8 ± 19 HU in group B (p = 0.75). The discriminating parameters of texture analysis were 10th (p = 0.03), 50th (p = 0.01), 90th percentiles (p = 0.04), S(0,1)SumAverg (p = 0.02), S(1,-1)SumOfSqs (p = 0.02), S(3,0)Contrast (p = 0.03), S(3,0)SumAverg (p = 0.02), S(4,0)SumAverg (p = 0.04), Horzl_RLNonUni (p = 0.02), and Vertl_LngREmph (p = 0.0005). Conclusions Texture analysis provides distinctive radiomic parameters of EF and TSF. EF and TSF had different radiomic features as the BMI varies.

The epicardial adipose tissue (EAT) is the visceral fat located between the myocardium and the pericardium. We aimed to perform a longitudinal evaluation of the epicardial adipose tissue using an advanced computer-assisted approach in a population of patients undergoing Cardiac CT (CCT) during statin administration, in relation to their body mass index (BMI). We retrospectively enrolled 95 patients [mean age 62 ± 10 years; 68 males (72%) and 27 females (28%)] undergoing CCT for suspected coronary artery disease during statin administration. CCT was performed at two subsequent time points. At the second CCT, EAT showed a mean density increase (-75.59 ± 7.0 HU vs. -78.18 ± 5.3 HU, p < 0.001) and a volume decrease (130 ± 54.3 cm3 vs.142.79 ± 56.9 cm3, p < 0.001). Concerning coronary artery EAT thickness, a reduction was found at the origin of the right coronary artery (13.26 ± 5.2 mm vs. 14.94 ± 5.8, p = 0.001) and interventricular artery (8.22 ± 3.7 mm vs. 9.13 ± 3.9 mm, p = 0.001). The quartile (Q) attenuation percentage (%) distribution of EAT changed at the second CCT. The EAT % distribution changed by the BMI in Q1 (p = 0.015), Q3 (p = 0.001) and Q4 (p = 0.043) at the second CCT, but the normal-BMI and over-weight/obese patients showed a similar response to statin therapy in terms of quartiles distribution changes. In conclusion, statins may determine significant changes in EAT quantitative and qualitative characteristics detected by CCT; the BMI influences the EAT composition, but statins determine a similar response in quartile distribution's variation, irrespective of the BMI.

Edge Analytics and Artificial Intelligence are important features of the current smart connected living community. In a society where people, homes, cities, and workplaces are simultaneously connected through various devices, primarily through mobile devices, a considerable amount of data is exchanged, and the processing and storage of these data are laborious and difficult tasks. Edge Analytics allows the collection and analysis of such data on mobile devices, such as smartphones and tablets, without involving any cloud-centred architecture that cannot guarantee real-time responsiveness. Meanwhile, Artificial Intelligence techniques can constitute a valid instrument to process data, limiting the computation time, and optimising decisional processes and predictions in several sectors, such as healthcare. Within this field, in this article, an approach able to evaluate the voice quality condition is proposed. A fully automatic algorithm, based on Deep Learning, classifies a voice as healthy or pathological by analysing spectrogram images extracted by means of the recording of vowel /a/, in compliance with the traditional medical protocol. A light Convolutional Neural Network is embedded in a mobile health application in order to provide an instrument capable of assessing voice disorders in a fast, easy, and portable way. Thus, a straightforward mobile device becomes a screening tool useful for the early diagnosis, monitoring, and treatment of voice disorders. The proposed approach has been tested on a broad set of voice samples, not limited to the most common voice diseases but including all the pathologies present in three different databases achieving F1-scores, over the testing set, equal to 80%, 90%, and 73%. Although the proposed network consists of a reduced number of layers, the results are very competitive compared to those of other "cutting edge" approaches constructed using more complex neural networks, and compared to the classic deep neural networks, for example, VGG-16 and ResNet-50.

The incidence of Head and Neck Squamous Cell Carcinoma (HNSCC) has been growing in the last few decades. Its diagnosis is usually performed through clinical evaluation and analyzing radiological images, then confirmed by histopathological examination, an invasive and time-consuming operation. The recent advances in the artificial intelligence field are leading to interesting results in the early diagnosis, personalized treatment and monitoring of HNSCC only by analyzing radiological images, without performing a tissue biopsy. The large amount of radiological images and the increasing interest in radiomics approaches can help to develop machine learning (ML) methods to support diagnosis. In this work, we propose an ML method based on the use of radiomics features, extracted from CT and PET images, to classify the disease in terms of pN-Stage, pT-Stage and Overall Stage. After the extraction of radiomics features, a selection step is performed to remove dataset redundancy. Finally, ML methods are employed to complete the classification task. Our pipeline is applied on the "Head-Neck-PET-CT" TCIA open-source dataset, considering a cohort of 201 patients from four different institutions. An AUC of 97%, 83% and 93% in terms of pN-Stage, pT-Stage and Overall Stage classification, respectively, is achieved. The obtained results are promising, showing the potential efficiency of the use of radiomics approaches in staging classification

Editorial of Special Issue on Analysis of 1D biomedical signals through AI based approaches for image processing

Breast cancer is the most commonly diagnosed cancer and registers the highest number of deaths for women. Advances indiagnostic activities combined with large-scale screening policies have significantly lowered themortality rates for breast cancerpatients. However, the manual inspection of tissue slides by pathologists is cumbersome, time-consuming, and is subject tosignificant inter- and intra-observer variability. Recently, the advent of whole-slide scanning systems has empowered the rapiddigitization of pathology slides, and enabled the development of Artificial Intelligence (AI)-assisted digital workflows. However, AItechniques, especially Deep Learning (DL), require a large amount of high-quality annotated data to learn from. Constructingsuch task-specific datasets poses several challenges, such as data-acquisition level constraints, time-consuming and expensiveannotations, and anonymization of patient information. In this paper, we introduce the BReAst Carcinoma Subtyping (BRACS)dataset, a large cohort of annotated Hematoxylin &amp; Eosin (H&amp;E)-stained images to advance AI development in the automaticcharacterization of breast lesions. BRACS contains 547Whole-Slide Images (WSIs), and 4539 Regions of Interest (RoIs) extractedfromtheWSIs. EachWSI, and respective RoIs, are annotated by the consensus of three board-certified pathologists into differentlesion categories. Specifically, BRACS includes three lesion types, i.e., benign, malignant and atypical, which are further subtypedinto seven categories. It is, to the best of our knowledge, the largest annotated dataset for breast cancer subtyping both atWSIandRoI-level. Further, by including the understudied atypical lesions, BRACS offers an unique opportunity for leveraging AI tobetter understand their characteristics. We encourage AI practitioners to develop and evaluate novel algorithms on the BRACSdataset to further breast cancer diagnosis and patient care.

Although the Convolutional Neural Networks (CNNs) have been widely adopted for the classification of histopathology images, one of the main drawbacks of CNNs is their inability to cope with gigapixel images. To deal with the high resolution of Whole Slide Image (WSI), many methods focus on patch processing which can result in improper representation if the patches are analyzed independently, losing the context information that is fundamental in digital pathology. In this study, the WSI is mapped into a compressed representation preserving the topological and morphological information relating to spatial correlations of neighboring patch features of the WSI. Such a representation is used to train a CNN to solve a classification task of breast histological images. The effectiveness of the suggested framework is demonstrated through experiments on Camelyon16 dataset. The results show the performance of our approach when three different ways to incorporate the spatial correlation in the tensor are used singly or in combination, highlighting that it is comparable with the state of the art.

Over the past couple of years, Transformers became increasingly popular within the deep learning community. Initially designed for Natural Language Processing tasks, Transformers were then tailored to fit to the Image Analysis field. The self-attention mechanism behind Transformers immediately appeared a promising, although computationally expensive, learning approach. However, Transformers do not adapt as well to tasks involving large images or small datasets. This propelled the exploration of hybrid CNN-Transformer models, which seemed to overcome those limitations, thus sparkling an increasing interest also in the field of medical imaging. Here, a hybrid approach is investigated for Pigment Signs (PS) segmentation in Fundus Images of patients suffering from Retinitis Pigmentosa, an eye disorder eventually leading to complete blindness. PS segmentation is a challenging task due to the high variability of their size, shape and colors and to the difficulty to distinguish between PS and blood vessels, which often overlap and display similar colors. To address those issues, we use the Group Transformer U-Net, a hybrid CNN-Transformer. We investigate the effects, on the learning process, of using different losses and choosing an appropriate parameter tuning. We compare the obtained performances with the classical U-Net architecture. Interestingly, although the results show margins for a consistent improvement, they do not suggest a clear superiority of the hybrid architecture. This evidence raises several questions, that we address here but also deserve to be further investigated, on how and when Transformers are really the best choice to address medical imaging tasks.

While cancer cases continue to increase and diagnosis, prognosis and treatment become more digital, AI-assisted cancer patient care, in particular in the pathology daily practice, remains scarce and rudimentary. In this chapter, we focus on reducing the gap between the AI technologies' outcomes and the way pathologists interpret the content of a histology image. We propose to leverage a semantic approach for both representing the relevant histology images and learning from them, thus mapping content to functionality and phenotype, that we call HistoCartography. We construct HierArchical Cell-to-Tissue (HACT) graphs to represent the content, leverage graph neural networks to learn from the HACT representations and respective graph explainers to indicate the image content that drives the AI-technologies' outputs. We further introduce a post-hoc graph explainer to quantitatively and qualitatively map the decision driving histology image content to measurable, pathologically understandable concepts. We test and validate the proposed approach by classifying seven breast carcinoma subtypes and demonstrate its power with respect to classification accuracy but moreover with respect to its ability to correlate to pathological knowledge and acceptance by domain experts, as validated by three independent pathologists from different institutions.

ancer diagnosis, prognosis, and therapy responsepredictions from tissue specimens highly depend on the phe-notype and topological distribution of constituting histologicalentities. Thus, adequate tissue representations for encodinghistological entities is imperative for computer aided cancerpatient care. To this end, several approaches have leveragedcell-graphs that encode cell morphology and cell organizationto denote the tissue information. These allow for utilizing graphtheory and machine learning to map tissue representationsto tissue functionality to help quantify their relationship.Though cellular information is crucial, it is incomplete aloneto comprehensively characterize complex tissue structure. Weherein treat the tissue as a hierarchical composition of multipletypes of histological entities from fine to coarse level, capturingmultivariate tissue information at multiple levels. We proposea novel multi-level hierarchical entity-graph representationof tissue specimens to model hierarchical compositions thatencode histological entities as well as their intra- and inter-entity level interactions. Subsequently, a hierarchical graphneural network is proposed to operate on the hierarchicalentity-graph representation to map the tissue structure to tissuefunctionality. Specifically, for input histology images we utilizewell-defined cells and tissue regions to build HierArchical Cell-to-Tissue (HACT) graph representations, and deviseHACT-Net, a message passing graph neural network, to classify suchHACTrepresentations. As part of this work, we introducethe BReAst Carcinoma Subtyping (BRACS) dataset, a largecohort of Haematoxylin &amp; Eosin stained breast tumor regions-of-interest, to evaluate and benchmark our proposed method-ology against pathologists and state-of-the-art computer-aideddiagnostic approaches. Through comparative assessment andablation studies, our proposed method is demonstrated to yieldsuperior classification results compared to alternative methodsas well as individual pathologists.

Retinitis Pigmentosa is a retinal disorder leading to a progressive visual field loss and eventually to complete blindness, but an early diagnosis could delay its progression through specific therapies. Retinitis pigmentosa is characterized by typical pigment signs that accumulate in the different regions of the retina. Pigment signs could be detected by a low-cost diagnosis tool, analyzing visual fundus retinal images and applying segmentation algorithms to annotate the pigments, so that, in a telemedicine scenario, the segmented images could be forwarded to an ophthalmologist for a rapid diagnosis. Deep learning approaches might be appropriate for this problem, but they have rarely been used to address it. However, pigment segmentation is a challenging task due to image resolution, small size of pigments and their proximity with blood vessels with which they share similar colors, and inter-patient widely changing image features. Very recently, transformer architectures, based on the self-attention paradigm, have emerged in the deep learning community as a powerful yet not completely explored tool to learn features directly from the data. Nonetheless they could not be directly exploited on small datasets, as they require a very large amount of data to learn meaningful features. To overcome the need for large training datasets, but also to reduce the high computation effort, hybrid architectures have been proposed, with the aim to combine the long-range relationship detection of transformers with the invariance and short-range detection properties of classical deep learning architectures. Here we investigate the performances of the Group Transformer U-Net, a hybrid approach for pigment segmentation on fundus images. This hybrid architecture modifies the classical U-Net structure introducing bottleneck multihead self-attention blocks between convolutional layers in both the contracting and expanding paths of the network. We compare the results obtained with this approach with the ones of the standard U-Net, and we describe how these results are affected when using different loss functions for the learning process, or strategies to address class imbalance.

The classification of images is of high importance in medicine. In this sense, Deep learning methodologies show excellent performance with regard to accuracy. The drawback of these methodologies is the fact that they are black boxes, so no explanation is given to users on the reasons underlying their choices. In the medical domain, this lack of transparency and information, typical of black box models, brings practitioners to raise concerns, and the result is a resistance to the use of deep learning tools. In order to overcome this problem, a different Machine Learning approach to image classification is used here that is based on interpretability concepts thanks to the use of an evolutionary algorithm. It relies on the application of two steps in succession. The first receives a set of images in the inut and performs image filtering on them so that a numerical data set is generated. The second is a classifier, the kernel of which is an evolutionary algorithm. This latter, at the same time, classifies and automatically extracts explicit knowledge as a set of IF-THEN rules. This method is investigated with respect to a data set of MRI brain imagery referring to Alzheimer's disease. Namely, a two-class data set (non-demented and moderate demented) and a three-class data set (non-demented, mild demented, and moderate demented) are extracted. The methodology shows good results in terms of accuracy (100% for the best run over the two-class problem and 91.49% for the best run over the three-class one), F_score (1.0000 and 0.9149, respectively), and Matthews Correlation Coefficient (1.0000 and 0.8763, respectively). To ascertain the quality of these results, they are contrasted against those from a wide set of well-known classifiers. The outcome of this comparison is that, in both problems, the methodology achieves the best results in terms of accuracy and F_score, whereas, for the Matthews Correlation Coefficient, it has the best result over the two-class problem and the second over the three-class one.

Within this paper a Grammatical Evolution algorithm is exploited to induce personalized and interpretable glucose forecasting models for diabetic patients based on the historical measurements of the glucose, the carbohydrates, and the injected insulin. A real-world data set of Type 1 diabetic patients is used to assess the induced models. The experimental trials show that the performance of extracted models is comparable with that obtained by other state-of-the-art techniques thatrequire a more significant computational effort.

The hypothesis that we intend to investigate here is that the extent of the impact of Covid-19 on a given country can be explained starting from a set of indicators and by using machine learning methodologies. The purpose of this chapter is not to find a way to solve the problem in an optimal way. Rather, we aim at performing a preliminary study to verify whether the aforementioned hypothesis is viable. Should it turn out so, we wish to get awareness both of which are the problems that must be solved in order to arrive at a (sub-)optimal solution, and of what are the possible limitations of the method. We firstly create a suitable data set of indicators starting from different sources available on the internet. Then, we apply onto it an evolutionary algorithm that is able to extract a set of IF-THEN decision rules allowing us to relate the values of the parameters for the different countries to the different levels of impact of Covid-19 on them.

In medical practice, all decisions, as for example the diagnosis based on the classification of images, must be made reliably and effectively. The possibility of having automatic tools helping doctors in performing these important decisions is highly welcome. Artificial Intelligence techniques, and in particular Deep Learning methods, have proven very effective on these tasks, with excellent performance in terms of classification accuracy. The problem with such methods is that they represent black boxes, so they do not provide users with an explanation of the reasons for their decisions. Confidence from medical experts in clinical decisions can increase if they receive from Artificial Intelligence tools interpretable output under the form of, e.g., explanations in natural language or visualized information. This way, the system outcome can be critically assessed by them, and they can evaluate the trustworthiness of the results. In this paper, we propose a new general-purpose method that relies on interpretability ideas. The approach is based on two successive steps, the former being a filtering scheme typically used in Content-Based Image Retrieval, whereas the latter is an evolutionary algorithm able to classify and, at the same time, automatically extract explicit knowledge under the form of a set of IF-THEN rules. This approach is tested on a set of chest X-ray images aiming at assessing the presence of COVID-19.

Diabetes is a heterogeneous group of diseases that share a common trait of elevated blood glucose levels. Insulin lowers this level by promoting glucose utilization, thus avoiding short- and long-term organ damage due to the elevated blood glucose level. A patient with diabetes uses an insulin pump to dose insulin. The pump uses a controller to compute and dose the correct amount of insulin to keep blood glucose levels in a safe range. Insulin-pump controller development is an ongoing process aiming at fully closed-loop control. Controllers entering the market must be evaluated for safety. We propose an evaluation method that exploits an FDA-approved diabetic patient simulator. The method evaluates a Cartesian product of individual insulin-pump parameters with a fine degree of granularity. As this is a computationally intensive task, the simulator executes on a distributed cluster. We identify safe and risky combinations of insulin-pump parameter settings by applying the binomial model and decision tree to this product. As a result, we obtain a tool for insulin-pump settings and controller safety assessment. In this paper, we demonstrate the tool with the Low-Glucose Suspend and OpenAPS controllers. For average +/- standard deviation, LGS and OpenAPS exhibited 1.7 +/- 0.6% and 3.2 +/- 1.8% of local extrema (i.e., good insulin-pump settings) out of all the entire Cartesian products, respectively. A continuous region around the best-discovered settings (i.e., the global extremum) of the insulin-pump settings spread across 4.0 +/- 1.1% and 4.1 +/- 1.3% of the Cartesian products, respectively.

This chapter presents an overview of the main Artificial Intelligence models used for time series data analysis, highlighting the main characteristics of each. The aim is to provide researchers with an panoramic that can guide them in choosing the most suitable technique for their studies.

Rationale and Objectives: To develop and validate a radiomic model, with radiomic features extracted from breast Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI), for the prediction of mass-enhancement lesion malignancy. Materials and Methods: 107 radiomic features were extracted from a manually annotated dataset of 111 patients, which was split into discovery and test sets. A feature calibration and pre-processing step was performed to find only robust non-redundant features. An in-depth discovery analysis was performed to define a predictive model: for this purpose, a Support Vector Machine (SVM) was trained in a nested 5-fold cross-validation scheme, by exploiting several unsupervised feature selection methods. The predictive model performance was evaluated in terms of Area Under the Receiver Operating Characteristic (AUROC), specificity, sensitivity, PPV and NPV. The test was performed on held-out set.Results: The model combining Unsupervised Discriminative Feature Selection (UDFS) and SVMs on average achieved the best performance on the blinded test set: AUROC = 0.725±0.091, sensitivity = 0.709±0.176, specificity = 0.741±0.114, PPV = 0.72±0.093, and NPV =0.75±0.114.Conclusion: In this study, we built a radiomic predictive model based on breast DCE-MRI, using only the strongest enhanced phase, with promising results in terms of accuracy and specificity in the differentiation of malignant from benign breast lesions.

The Special Issue “Artificial Intelligence Applied to Medical Imaging and Computational Biology” of the Applied Sciences Journal has been curated from February 2021 to May 2022, which covered the state-of-the-art and novel algorithms and applications of Artificial Intelligence methods for biomedical data analysis, ranging from classic Machine Learning to Deep Learning.

Unsupervised segmentation techniques, which do not require labeled data for training and can be more easily integrated into the clinical routine, represent a valid solution especially from a clinical feasibility perspective. Indeed, large-scale annotated datasets are not always available, under-mining their immediate implementation and use in the clinic. Breast cancer is the most common cause of cancer death in women worldwide. In this study, breast lesion delineation in Dynamic Contrast Enhanced MRI (DCE-MRI) series was addressed by means of four popular unsupervised segmentation approaches: Split-and-Merge combined with Region Growing (SMRG), k-means, Fuzzy C-Means (FCM), and spatial FCM (sFCM). They represent well-established pattern recognition techniques that are still widely used in clinical research. Starting from the basic versions of these segmentation approaches, during our analysis, we identified the shortcomings of each of them, proposing improved versions, as well as developing ad hoc pre-and post-processing steps. The obtained experimental results, in terms of area-based--namely, Dice Index (DI), Jaccard Index (JI), Sensitivity, Specificity, False Positive Ratio (FPR), False Negative Ratio (FNR)--and distance-based metrics--Mean Absolute Distance (MAD), Maximum Distance (MaxD), Hausdorff Distance (HD)--encourage the use of unsupervised machine learning techniques in medical image segmentation. In particular, fuzzy clustering approaches (namely, FCM and sFCM) achieved the best performance. In fact, for area-based metrics, they obtained DI = 78.23% ± 6.50 (sFCM), JI = 65.90% ± 8.14 (sFCM), sensitivity = 77.84% ± 8.72 (FCM), specificity = 87.10% ± 8.24 (sFCM), FPR = 0.14 ± 0.12 (sFCM), and FNR = 0.22 ± 0.09 (sFCM). Concerning distance-based metrics, they obtained MAD = 1.37 ± 0.90 (sFCM), MaxD = 4.04 ± 2.87 (sFCM), and HD = 2.21 ± 0.43 (FCM). These experimental findings suggest that further research would be useful for advanced fuzzy logic techniques specifically tailored to medical image segmentation.

Machine learning models based on radiomic features allow us to obtain biomarkers capable of modeling the disease and able to support the clinical routine. Recent studies showed that it is fundamental that the computed features are robust and reproducible. Although several initiatives to standardize the definition and extraction process of biomarkers are ongoing, there is lack of comprehensive guidelines. Therefore, no standardized procedures are available for ROI selection, feature extraction and processing, with the risk of undermining the effective use of radiomic models in clinical routine. In this study, we aim to assess the impact that the different segmentation methods and the quantization level (defined by means of the number of bins used in the features extraction phase) may have on the robustness of the radiomic features. In particular, the robustness of texture features extracted by PyRadiomics, and belonging to five categories---GLCM, GLRLM, GLSZM, GLDM, NGTDM---was evaluated using the Intra-class Correlation Coefficient (ICC) and mean differences between segmentation raters. In addition to the robustness each single feature, an overall index for each feature category was quantified. The analysis shows that the level of quantization (i.e., `bincount' parameter) plays a key role in defining robust features: in fact, in our study focused on a Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) dataset of 111 breast masses, sets with cardinality varying between 34 and 43 robust features, were obtained with `binCount' values equal to 256 and 32, respectively. Moreover, both manual segmentation methods demonstrated good reliability and agreement, instead automated segmentation achieved lower ICC values. Considering the dependence on the quantization level, taking into account only the \emph{intersection subset} among all the values of `binCount' could be the best selection strategy. Among radiomic feature categories, GLCM, GLRLM, and GLDM showed the best overall robustness as the segmentation method varies.

Multiparametric Magnetic Resonance Imaging (MRI) is the most sensitive imaging modality for breast cancer detection and is increasingly playing a key role in lesion characterization. In this context, accurate and reliable quantification of the shape and extent of breast cancer is crucial in clinical research environments. Since conventional lesion delineation procedures are still mostly manual, automated segmentation approaches can improve this time-consuming and operator-dependent task by annotating the regions of interest in a reproducible manner. In this work, a semi-automated and interactive approach based on the spatial Fuzzy C-Means (sFCM) algorithm is proposed, used to segment masses on dynamic contrast-enhanced (DCE) MRI of the breast. Our method was compared against existing approaches based on classic image processing, namely (i) Otsu's method for thresholding-based segmentation, and (ii) the traditional FCM algorithm. A further comparison was performed against state-of-the-art Convolutional Neural Networks (CNNs) for medical image segmentation, namely SegNet and U-Net, in a 5-fold cross-validation scheme. The results showed the validity of the proposed approach, by significantly outperforming the competing methods in terms of the Dice similarity coefficient (84.47±4.75). Moreover, a Pearson's coefficient of r=0.993 showed a high correlation between segmented volume and the gold standard provided by clinicians. Overall, the proposed method was confirmed to outperform the competing literature methods. The proposed computer-assisted approach could be deployed into clinical research environments by providing a reliable tool for volumetric and radiomics analyses.

Convolutional Neural Networks (CNNs) have proven to be one of the most powerful tools for solving complex problems in the field of pattern recognition and image analysis, even if serious challenges remain. Indeed, one of the main drawbacks of CNNs is their inability to cope with very high-resolution images. In areas other than digital pathology, image resizing is often the simplest and most effective solution. However, histopathological images not only show a very high resolution, but also contain a lot of information at the detail level, making this strategy completely ineffective. Other approaches partition theimage into small patches and analyze them independently, losing the context information that is fundamental in digital pathology. In this paper, we present a method based on a compressed representation of the Whole Slide Image (WSI), by building a 3D tensor, that preserves the topological and morphological information relating to the proximity relationships between the patches of the WSI. Tensors are used to train a CNN to solve a binary classification task. This technique has been evaluated for the analysis of gigapixel Hematoxylin and Eosin (H&amp;E) histological images with the aim of supporting the diagnosis of breast cancer. Several experiments have been performed on the Camelyon16 dataset by generating different types of 3D tensors. The results of the proposed approach on the breast cancer classification task have been compared with some state-of-the-art approaches.

The BReAst Carcinoma Subtyping (BRACS) is a new dataset of hematoxylin and eosin (H&E) histopathological images of breast carcinoma. BRACS has been built on the basis of an Agreement between IRCCS Fondazione Pascale, Institute for High Performance Computing and Networking (ICAR) of National Research Council (CNR), and IBM Research-Zurich for the "Development of methodologies and tools for the identification of atypical tumors in breast cancer pathology through the automatic analysis of histological images". This dataset offers a platform for researchers to compare strategies and algorithms for automated detection/classification of breast tumors in H&E stained tissue samples collected by mastectomy or biopsy. BRACS differs from most of the public breast cancer image datasets since it includes images representing atypical lesions. An early diagnosis of these atypical lesions could prevent the worsening into malignant cancer. In details, BRACS contains images characterized by the following kind of lesions: Pathological Benign (PB), Usual Ductal Hyperplasia (UDH), Flat Epithelial Atypia (FEA), Atypical Ductal Hyperplasia (ADH), Ductal Carcinoma in Situ (DCIS) and Invasive Carcinoma (IC). Also images representing Normal (N) tissue samples, i.e. glandular tissue samples without lesions, are included into BRACS.

Although CNNs are widely considered as the state-of-the-art models in various applications of image analysis, one of the main challenges still open is the training of a CNN on high resolution images. Different strategies have been proposed involving either a rescaling of the image or an individual processing of parts of the image. Such strategies cannot be applied to images, such as gigapixel histopathological images, for which a high reduction in resolution inherently effects a loss of discriminative information, and in respect of which the analysis of single parts of the image suffers from a lack of global information or implies a high workload in terms of annotating the training images in such a way as to select significant parts. We propose a method for the analysis of gigapixel histopathological images solely by using weak image-level labels. In particular, two analysis tasks are taken into account: a binary classification and a prediction of the tumor proliferation score. Our method is based on a CNN structure consisting of a compressing path and a learning path. In the compressing path, the gigapixel image is packed into a grid-based feature map by using a residual network devoted to the feature extraction of each patch into which the image has been divided. In the learning path, attention modules are applied to the grid-based feature map, taking into account spatial correlations of neighboring patch features to find regions of interest, which are then used for the final whole slide analysis. Our method integrates both global and local information, is flexible with regard to the size of the input images and only requires weak image-level labels. Comparisons with different methods of the state-of-the-art on two well known datasets, Camelyon16 and TUPAC16, have been made to confirm the validity of the proposed model.

Cardiovascular diseases cause the death of almost 18 million people each year. Heart failure takes place when the heart does not pump a sufficient amount of blood to the body and is one of the most common causes of death. Healthcare technologies and Artificial Intelligence algorithms constitute promising solutions useful not only for the monitoring of people's health but also in order to tell which subjects are more prone to this problem, which is information of paramount relevance to save their lives. The goal of this paper is to understand the predictability of mortality of subjects suffering from left ventricular systolic dysfunction who previously experienced heart failures. To perform this important study, a publicly-available data set is considered that contains thirteen pieces of clinical, body, and lifestyle information about 299 subjects. In tackling this data set, not only do we wish to perform classification with reference to subjects' survival/death, but we also wish to automatically extract explainable knowledge about the reasons for the classification proposed. To this aim, we use DEREx, an Artificial Intelligence-based tool that relies on Evolutionary Algorithms and provides users with an easy-to-understand set of IF-THEN rules containing data set parameters. In this way, it performs the selection of the parameters that are the most relevant for the purpose of classification. We have run our experiments following a sound protocol established in the scientific literature for this data set. Our findings show that, apart from automatically obtaining easily interpretable knowledge, DEREx achieves better results in terms of widely-used quality indices as Matthews Correlation Coefficient, accuracy, and F score.

Several compartmental models have been designed in epidemiology to simplify the mathematical modeling of infectious diseases, so as to describe their spreading in a population of individuals. Among them, we will make use here of the SEIAR that expands the basic SIR model and the SEIR one. The choice of SEIAR model is due to the fact that we wish to estimate here the spreading of the coronavirus COVID-19 in Italy. In fact, several papers have stressed the issue that for this pandemic, the number of asymptomatic infectious subjects is very high. Given that asymptomatic subjects are obviously not contained in official figures, their presence causes a much wider and longer spread of this disease, with more infectious people. Moreover, an important remark on the use of the SEIAR model is that the basic reproduction number R0 it computes is much higher than that provided by the use of SIR and SEIR models.

University of Messina, Messina, University of Messina, Messina, Italy; ICAR-CNR, Naples, ICAR-CNR, Naples, Italy; ICAR-CNR, Naples, ICAR-CNR, Naples, Italy; University of Warwick, Coventry, University of Warwick, Coventry, United Kingdom

The quality of life of diabetic patients can be enhanced by devising an artificial pancreas endowed with a personalized control algorithm able to regulate the insulin dosage.A fundamental step in the building of this device is to conceive an efficient algorithm for forecasting future glucose levels.Within this paper, an evolutionary-based strategy, i.e., a Grammatical Evolution algorithm, is devised to deduce a personalized regression model able to estimate future blood glucose values onthe basis of the past glucose measurements, and the knowledge of the food intake, and of the basal and injected insulin levels.The aim is to discover models that are not only interpretable but also with low complexity to be used within a control algorithm that is the main element of the artificial pancreas. A real-worlddatabase composed by patients suffering from Type 1 diabetes has been employed to evaluate the proposed evolutionary automatic procedure.

Diabetes mellitus is a lifelong disease in which either the pancreas fails to produce insulin or the produced amount is insufficient to control blood sugar levels. A way to tackle this malfunctioning is to devise an artificial pancreas endowed with a personalized control algorithm able to regulate the insulin dosage. A crucial step in realizing such a device is to effectively forecast future glucose levels starting from past glucose values, the knowledge of the food intake, and of the basal and the injected insulin. The increasing availability of medical diabetes data sets is providing unprecedented opportunities to identify correlations inside these data even harnessing innovative investigation methods, such as deep learning. As an alternative to the deep learning methods successfully used as forecasting models, we exploit a neuroevolution algorithm to model and predict future personalized blood glucose levels. The discovered subjective regression model can represent the control algorithm of an artificial pancreas. This model is assessed through experiments performed on a real-world database containing data of six patients suffering from Type 1 diabetes. To further evaluate the effectiveness of the predictions derived from the proposed approach, the results are also compared against those obtained by other state-of-the-art recently proposed methods.

The absence of the rapid eye movement (REM) phase during sleep can have negative consequences, as, e.g., anxiety, increase in appetite, irritability, while, on the other hand, it can help in improving some kinds of depression. The goal of the research described in this chapter consists in the identification of the different sleep phases a subject is experiencing by using heart rate variability (HRV) values. These are computed starting from the signals gathered from electrocardiogram (ECG) sensors placed on the subject. To this aim, the publicly available Sleep Heart Health Study (SHHS) data set is taken into account, which contains both types of information. Several machine learning classification algorithms are tested on this data set, and their performance is compared in terms of F1-score value, as SHHS is highly unbalanced. Once the most suitable classification algorithm is found, it can be firstly trained offline on the problem and then used online in an IoT-based fully automated e-health system. In this latter, sensors gather, in real time, ECG signals from a sleeping subject, send them to a device where data is processed, HRV values are computed, sleep phase identification takes place, and medical personnel, close or not to the subject, are immediately informed of the subject's sleeping phases.

Structural and metabolic imaging are fundamental for diagnosis, treatment and follow-up in oncology. Beyond the well-established diagnostic imaging applications, ultrasounds are currently emerging in the clinical practice as a noninvasive technology for therapy. Indeed, the sound waves can be used to increase the temperature inside the target solid tumors, leading to apoptosis or necrosis of neoplastic tissues. The Magnetic resonance-guided focused ultrasound surgery (MRgFUS) technology represents a valid application of this ultrasound property, mainly used in oncology and neurology. In this paper; patient safety during MRgFUS treatments was investigated by a series of experiments in a tissue-mimicking phantom and performing ex vivo skin samples, to promptly identify unwanted temperature rises. The acquired MR images, used to evaluate the temperature in the treated areas, were analyzed to compare classical proton resonance frequency (PRF) shift techniques and referenceless thermometry methods to accurately assess the temperature variations. We exploited radial basis function (RBF) neural networks for referenceless thermometry and compared the results against interferometric optical fiber measurements. The experimental measurements were obtained using a set of interferometric optical fibers aimed at quantifying temperature variations directly in the sonication areas. The temperature increases during the treatment were not accurately detected by MRI-based referenceless thermometry methods, and more sensitive measurement systems, such as optical fibers, would be required. In-depth studies about these aspects are needed to monitor temperature and improve safety during MRgFUS treatments.

The recent developments of information technologies, and the consequent need for access to distributed services and resources, require robust and reliable authentication systems. Biometric systems can guarantee high levels of security and multimodal techniques, which combine two or more biometric traits, warranting constraints that are more stringent during the access phases. This work proposes a novel multimodal biometric system based on iris and retina combination in the spatial domain. The proposed solution follows the alignment and recognition approach commonly adopted in computational linguistics and bioinformatics; in particular, features are extracted separately for iris and retina, and the fusion is obtained relying upon the comparison score via the Levenshtein distance. We evaluated our approach by testing several combinations of publicly available biometric databases, namely one for retina images and three for iris images. To provide comprehensive results, detection error trade-off-based metrics, as well as statistical analyses for assessing the authentication performance, were considered. The best achieved False Acceptation Rate and False Rejection Rate indices were and 3.33%, respectively, for the multimodal retina-iris biometric approach that overall outperformed the unimodal systems. These results draw the potential of the proposed approach as a multimodal authentication framework using multiple static biometric traits.

Recent advances in Quantum Machine Learning (QML) have provided benefits to several computational processes, drastically reducing the time complexity. Another approach of combining quantum information theory with machine learning--without involving quantum computers--is known as Quantum-inspired Machine Learning (QiML), which exploits the expressive power of the quantum language to increase the accuracy of the process (rather than reducing the time complexity). In this work, we propose a large-scale experiment based on the application of a binary classifier inspired by quantum information theory to the biomedical imaging context in clonogenic assay evaluation to identify the most discriminative feature, allowing us to enhance cell colony segmentation. This innovative approach offers a two-fold result: (1) among the extracted and analyzed image features, homogeneity is shown to be a relevant feature in detecting challenging cell colonies; and (2) the proposed quantum-inspired classifier is a novel and outstanding methodology, compared to conventional machine learning classifiers, for the evaluation of clonogenic assays.

Biometric classification plays a key role in fingerprint characterization, especially in the identification process. In fact, reducing the number of comparisons in biometric recognition systems is essential when dealing with large-scale databases. The classification of fingerprints aims to achieve this target by splitting fingerprints into different categories. The general approach of fingerprint classification requires pre-processing techniques that are usually computationally expensive. Deep Learning is emerging as the leading field that has been successfully applied to many areas, such as image processing. This work shows the performance of pre-trained Convolutional Neural Networks (CNNs), tested on two fingerprint databases--namely, PolyU and NIST--and comparisons to other results presented in the literature in order to establish the type of classification that allows us to obtain the best performance in terms of precision and model efficiency, among approaches under examination, namely: AlexNet, GoogLeNet, and ResNet. We present the first study that extensively compares the most used CNN architectures by classifying the fingerprints into four, five, and eight classes. From the experimental results, the best performance was obtained in the classification of the PolyU database by all the tested CNN architectures due to the higher quality of its samples. To confirm the reliability of our study and the results obtained, a statistical analysis based on the McNemar test was performed.

We introduce a machine learning-based analysis to predict the immunohistochemical (IHC) labeling index for the cell proliferation marker Ki67/MIB1 on cancer tissues based on morphometrical features extracted from hematoxylin and eosin (H&amp;E)-stained formalin-fixed, paraffin-embedded (FFPE) tumor tissue samples. We provided a proof-of-concept prediction of Ki67/MIB1 IHC positivity of cancer cells through the definition and quantitation of single nuclear features. In the first instance, we set our digital framework on Ki67/MIB1-stained OSCC tissue sample whole slide images, using QuPath as a working platform and its integrated algorithms, we built a classifier in order to distinguish tumor and stroma classes and, within them, Ki67-positive and Ki67-negative cells; then we sorted out morphometric features of tumor cells related to their Ki67 IHC status. Among the evaluated features, Nuclear Hematoxylin Mean Optical Density (NHMOD) resulted as the best one to distinguish Ki67/MIB1 positive from negative cells. We confirmed our findings in a single-cell level analysis of H&amp;E staining on Ki67-immunostained/H&amp;E-decolored tissue samples. We finally tested our digital framework on a case series of Oral Squamous cell carcinomas (OSCC), arranged in tissue microarrays; we selected two consecutive sections of each OSCC FFPE TMA block, respectively stained with H&amp;E and immuno-stained for Ki67/MIB1. We automatically detected tumor cells in H&amp;E slides, and we generated a "false color map" (FCM), based on NHMOD through the QuPath measurements map tool. FCM nearly coincided with the actual immunohistochemical result, allowing the prediction of Ki67/MIB1 positive cells in a direct visual fashion. Our proposed approach provides the pathologist with a fast method to identify the proliferating compartment of the tumor through a quantitative assessment of a nuclear feature on H&amp;E slides, readily appreciable by visual inspection. Although this technique needs to be fine-tuned and tested on larger series of tumors, the digital analysis approach appears as a promising tool to quickly forecast the tumor's proliferation fraction directly on H&amp;E routinely stained digital section.

Cancer diagnosis, prognosis, and therapeutic response prediction are heavily influenced by the relationship between the histopathological structures and the function of the tissue. Recent approaches acknowledging the structurefunction relationship, have linked the structural and spatial patterns of cell organization in tissue via cell-graphs to tumor grades. Though cell organization is imperative, it is insufficient to entirely represent the histopathological structure. We propose a novel hierarchical cell-to-tissue-graph (HACT) representation to improve the structural depiction of the tissue. It consists of a low-level cell-graph, capturing cell morphology and interactions, a high-level tissue-graph, capturing morphology and spatial distribution of tissue parts, and cells-totissue hierarchies, encoding the relative spatial distribution of the cells with respect to the tissue distribution. Further, a hierarchical graph neural network (HACT-Net) is proposed to efficiently map the HACT representations to histopathological breast cancer subtypes. We assess the methodology on a large set of annotated tissue regions of interest from H&amp;E stained breast carcinoma whole-slides. Upon evaluation, the proposed method outperformed recent convolutional neural network and graph neural network approaches for breast cancer multi-class subtyping. The proposed entity-based topological analysis is more inline with the pathological diagnostic procedure of the tissue. It provides more command over the tissue modelling, therefore encourages the further inclusion of pathological priors into task-specific tissue representation

The aim of this paper consists in the application of a recent epidemiological model, namely SEIR with Social Distancing(SEIR-SD), extended here through the definition of a social distancing function varying over time, to assess the situation relatedto the spreading of the coronavirus Covid-19 in Italy and in two of its most important regions, i.e., Lombardy and Campania. Toprofitably use this model, the most suitable values of its parameters must be found. The estimation of the SEIR-SD modelparameters takes place here through the use of Differential Evolution, a heuristic optimization technique. In this way, we areable to evaluate for each of the three above-mentioned scenarios the daily number of infectious cases from today until theend of virus spreading, the day(s) in which this number will be at its highest peak, and the day in which the infected cases willbecome very close to zero.

Automatic detection of falls is extremely important, especially in the remote monitoring of elderly people, and will become more and more critical in the future, given the constant increase in the number of older adults. Within this framework, this paper deals with the task of evaluating several artificial intelligence techniques to automatically distinguish between different activities of daily living (ADLs) and different types of falls. To do this, UniMiB SHAR, a publicly available data set containing instances of nine different ADLs and of eight kinds of falls, is considered. We take into account five different classes of classification algorithms, namely tree-based, discriminant-based, support vector machines, K-nearest neighbors, and ensemble mechanisms, and we consider several representatives for each of these classes. These are all the classes contained in the Classification Learner app contained in MATLAB, which serves as the computational basis for our experiments. As a result, we apply 22 different classification algorithms coming from artificial intelligence under a fivefold cross-validation learning strategy, with the aim to individuate which the most suitable is for this data set. The numerical results show that the algorithm with the highest classification accuracy is the ensemble based on subspace as the ensemble method and on KNN as learner type. This shows an accuracy equal to 86.0%. Its results are better than those in the other papers in the literature that face this specific 17-class problem.

One of the most important physiological parameters of the cardiovascular circulatory system is Blood Pressure. Several diseases are related to long-term abnormal blood pressure, i.e., hypertension; therefore, the early detection and assessment of this condition are crucial. The identification of hypertension, and, even more the evaluation of its risk stratification, by using wearable monitoring devices are now more realistic thanks to the advancements in Internet of Things, the improvements of digital sensors that are becoming more and more miniaturized, and the development of new signal processing and machine learning algorithms. In this scenario, a suitable biomedical signal is represented by the PhotoPlethysmoGraphy (PPG) signal. It can be acquired by using a simple, cheap, and wearable device, and can be used to evaluate several aspects of the cardiovascular system, e.g., the detection of abnormal heart rate, respiration rate, blood pressure, oxygen saturation, and so on. In this paper, we take into account the Cuff-Less Blood Pressure Estimation Data Set that contains, among others, PPG signals coming from a set of subjects, as well as the Blood Pressure values of the latter that is the hypertension level. Our aim is to investigate whether or not machine learning methods applied to these PPG signals can provide better results for the non-invasive classification and evaluation of subjects' hypertension levels. To this aim, we have availed ourselves of a wide set of machine learning algorithms, based on different learning mechanisms, and have compared their results in terms of the effectiveness of the classification obtained.

Natural phenomena and mechanisms have always intrigued humans, inspiring the design of effective solutions for real-world problems. Indeed, fascinating processes occur in nature, giving rise to an ever-increasing scientific interest. In everyday life, the amount of heterogeneous biomedical data is increasing more and more thanks to the advances in image acquisition modalities and high-throughput technologies. The automated analysis of these large-scale datasets creates new compelling challenges for data-driven and model-based computational methods. The application of intelligent algorithms, which mimic natural phenomena, is emerging as an effective paradigm for tackling complex problems, by considering the unique challenges and opportunities pertaining to biomedical images. Therefore, the principal contribution of computer science research in life sciences concerns the proper combination of diverse and heterogeneous datasets-i.e., medical imaging modalities (considering also radiomics approaches), Electronic Health Record engines, multi-omics studies, and real-time monitoring-to provide a comprehensive clinical knowledge. In this paper, the state-of-the-art of nature-inspired medical image analysis methods is surveyed, aiming at establishing a common platform for beneficial exchanges among computer scientists and clinicians. In particular, this review focuses on the main natureinspired computational techniques applied to medical image analysis tasks, namely: physical processes, bio-inspired mathematical models, Evolutionary Computation, Swarm Intelligence, and neural computation. These frameworks, tightly coupled with Clinical Decision Support Systems, can be suitably applied to every phase of the clinical workflow. We show that the proper combination of quantitative imaging and healthcare informatics enables an in-depth understanding of molecular processes that can guide towards personalised patient care.

Prostate cancer is the most common cancer among US men. However, prostate imaging is still challenging despite the advances in multi-parametric magnetic resonance imaging (MRI), which provides both morphologic and functional information pertaining to the pathological regions. Along with whole prostate gland segmentation, distinguishing between the central gland (CG) and peripheral zone (PZ) can guide toward differential diagnosis, since the frequency and severity of tumors differ in these regions; however, their boundary is often weak and fuzzy. This work presents a preliminary study on deep learning to automatically delineate the CG and PZ, aiming at evaluating the generalization ability of convolutional neural networks (CNNs) on two multi-centric MRI prostate datasets. Especially, we compared three CNN-based architectures: SegNet, U-Net, and pix2pix. In such a context, the segmentation performances achieved with/without pre-training were compared in 4-fold cross-validation. In general, U-Net outperforms the other methods, especially when training and testing are performed on multiple datasets.

A clonogenic assay is a biological technique for calculating the Surviving Fraction (SF) that quantifies the anti-proliferative effect of treatments on cell cultures: this evaluation is often performed via manual counting of cell colony-forming units. Unfortunately, this procedure is error-prone and strongly affected by operator dependence. Besides, conventional assessment does not deal with the colony size, which is generally correlated with the delivered radiation dose or administered cytotoxic agent. Relying upon the direct proportional relationship between the Area Covered by Colony (ACC) and the colony count and size, along with the growth rate, we propose MF2C3, a novel computational method leveraging spatial Fuzzy C-Means clustering on multiple local features (i.e., entropy and standard deviation extracted from the input color images acquired by a general-purpose flat-bed scanner) for ACC-based SF quantification, by considering only the covering percentage. To evaluate the accuracy of the proposed fully automatic approach, we compared the SFs obtained by MF2C3 against the conventional counting procedure on four different cell lines. The achieved results revealed a high correlation with the ground-truth measurements based on colony counting, by outperforming our previously validated method using local thresholding on L*u*v* color well images. In conclusion, the proposed multi-feature approach, which inherently leverages the concept of symmetry in the pixel local distributions, might be reliably used in biological studies.

Advances in microscopy imaging technologies have enabled the visualization of live-cell dynamic processes using time-lapse microscopy imaging. However, modern methods exhibit several limitations related to the training phases and to time constraints, hindering their application in the laboratory practice. In this work, we present a novel method, named Automated Cell Detection and Counting (ACDC), designed for activity detection of fluorescent labeled cell nuclei in time-lapse microscopy. ACDC overcomes the limitations of the literature methods, by first applying bilateral filtering on the original image to smooth the input cell images while preserving edge sharpness, and then by exploiting the watershed transform and morphological filtering. Moreover, ACDC represents a feasible solution for the laboratory practice, as it can leverage multi-core architectures in computer clusters to efficiently handle large-scale imaging datasets. Indeed, our Parent-Workers implementation of ACDC allows to obtain up to a 3.7× speed-up compared to the sequential counterpart. ACDC was tested on two distinct cell imaging datasets to assess its accuracy and effectiveness on images with different characteristics. We achieved an accurate cell-count and nuclei segmentation without relying on large-scale annotated datasets, a result confirmed by the average Dice Similarity Coefficients of 76.84 and 88.64 and the Pearson coefficients of 0.99 and 0.96, calculated against the manual cell counting, on the two tested datasets.

Accurately identifying and categorizing cancer structures/sub-types in histological images is an important clinical task involving a considerable workload and a specific subspecialty of pathologists. Digitizing pathology is a current trend that provides large amounts of visual data allowing a faster and more precise diagnosis through the development of automatic image analysis techniques. Recent studies have shown promising results for the automatic analysis of cancer tissue by using deep learning strategies that automatically extract and organize the discriminative information from the data. This paper explores deep learning methods for the automatic analysis of Hematoxylin and Eosin stained histological images of breast cancer and lymphoma. In particular, a deep learning approach is proposed for two different use cases: the detection of invasive ductal carcinoma in breast histological images and the classification of lymphoma sub-types. Both use cases have been addressed by adopting a Residual Convolutional Neural Network which is part of a Convolutional Autoencoder Network (i.e. FusionNet). The performances have been evaluated on public datasets of digital histological images and have been compared with those obtained by using different deep neural networks (UNet and ResNet). Additionally, comparisons with the state of the art have been considered, in accordance with different deep learning approaches. The experimental results show an improvement of 5:06% in F-measure score for the detection task, and an improvement of 1:09% in the accuracy measure for the classification task.

New technological advances in automated microscopy have given rise to large volumes of data, which have made human-based analysis infeasible, heightening the need for automatic systems for high-throughput microscopy applications. In particular, in the field of fluorescence microscopy, automatic tools for image analysis are making an essential contribution in order to increase the statistical power of the cell analysis process. The development of these automatic systems is a difficult task due to both the diversification of the staining patterns and the local variability of the images. In this paper, we present an unsupervised approach for automatic cell segmentation and counting, namely CSC, in high-throughput microscopy images. The segmentation is performed by dividing the whole image into square patches that undergo a gray level clustering followed by an adaptive thresholding. Subsequently, the cell labeling is obtained by detecting the centers of the cells, using both distance transform and curvature analysis, and by applying a region growing process. The advantages of CSC are manifold. The foreground detection process works on gray levels rather than on individual pixels, so it proves to be very efficient. Moreover, the combination of distance transform and curvature analysis makes the counting process very robust to clustered cells. A further strength of the CSC method is the limited number of parameters that must be tuned. Indeed, two different versions of the method have been considered, CSC-7 and CSC-3, depending on the number of parameters to be tuned. The CSC method has been tested on several publicly available image datasets of real and synthetic images. Results in terms of standard metrics and spatially-aware measures show that CSC outperforms the current state of the art techniques.

Breast cancer is the most common invasive cancer in women, aecting more than 10%of women worldwide. Microscopic analysis of a biopsy remains one of the most importantmethods to diagnose the type of breast cancer. This requires specialized analysisby pathologists, in a task that i) is highly time- and cost-consuming and ii) oftenleads to nonconsensual results. The relevance and potential of automatic classificationalgorithms using hematoxylin-eosin stained histopathological images has alreadybeen demonstrated, but the reported results are still sub-optimal for clinical use. Withthe goal of advancing the state-of-the-art in automatic classification, the Grand Challengeon BreAst Cancer Histology images (BACH) was organized in conjunction withthe 15th International Conference on Image Analysis and Recognition (ICIAR 2018).BACH aimed at the classification and localization of clinically relevant histopathologicalclasses in microscopy and whole-slide images from a large annotated dataset,specifically compiled and made publicly available for the challenge. Following a positiveresponse from the scientific community, a total of 64 submissions, out of 677registrations, eectively entered the competition. The submitted algorithms improvedthe state-of-the-art in automatic classification of breast cancer with microscopy imagesto an accuracy of 87%. Convolutional neuronal networks were the most successfulmethodology in the BACH challenge. Detailed analysis of the collective results allowedthe identification of remaining challenges in the field and recommendations for futuredevelopments. The BACH dataset remains publicly available as to promote further improvementsto the field of automatic classification in digital pathology.

The adoption of Deep Learning (DL) algorithms into the practice of ophthalmology could play an important role in screening and diagnosis of eye diseases in the coming years. In particular, DL tools interpreting ocular data derived from low-cost devices, as a fundus camera, could support massive screening also in resource limited countries. This paper explores a fully automatic method supporting the diagnosis of the Retinitis Pigmentosa by means of the segmentation of pigment signs in retinal fundus images. The proposed approach relies on an U-Net based deep convolutional network. At the present, this is the first approach for pigment signs segmentation in retinal fundus images that is not dependent on hand-crafted features, but automatically learns a hierarchy of increasingly complex features directly from data.We assess the performance by training the model on the public dataset RIPS and comparisons with the state of the art have been considered in accordance with approaches working on the same dataset. The experimental results show an improvement of 15% in F-measure score.

Health 4.0 can provide effective ways to improve the health status of subjects by taking advantage of Cyber-Physical Systems and Internet of Things technologies for the solution of health care problems. One of these is represented by suitably estimating blood pressure values of subjects in a continuous, real-time and non-invasive way. To address it, we propose an approach only requiring a photoplethysmography sensor and a mobile/desktop device. The approach avails itself of Genetic Programming to automatically find an explicit relationship between blood pressure values and photoplethysmography ones. This relationship is tested on a set of eleven subjects and compared against other regression methods, and turns out to be better. Namely, the Root Mean Square Error values are equal to 8.49 and 6.66 for the systolic and the diastolic BP values, respectively. Those for the relative error, instead, are equal to 5.55% for the systolic and 6.59% for the diastolic values.

Deep Neural Networks (DNNs) may be very effective for the classification over highly-sized data sets, especially in the medical domain, where the recognition of the occurrence of a specific event related to a disease is of high importance. Unfortunately, DNNs suffer from the drawback that a good set of values for their configuration hyper-parameters must be found. Currently, this is done through the use of either trial-and-error methods or sampling-based ones. In this paper we propose a new approach to find the most suitable structure for a DNN used for a classification problem in terms of achievement of the highest classification accuracy. This approach is based on a distributed version of Differential Evolution (DE), a variety of an Evolutionary Algorithm. To evaluate the approach, in this paper we investigate this issue with reference to Obstructive Sleep Apnea (OSA). OSA is an important medical problem consisting of episodes taking place during night in which a subject stops breathing due to a constriction of the upper airways. This deteriorates the quality of life and may have dangerous, and even lethal, consequences on both short and long term. An accurate classification is a very crucial step for the OSA treatment, because understanding automatically that a subject is experiencing such an episode may be decisive if prompt medical action is needed. In our experiments, classification takes place on a data set in which each item contains the values of 17 Heart Rate Variability parameters, extracted from ElectroCardiography signals, and the annotation of OSA events. We have extracted this data set from the real-world Sleep Heart Health Study database. The results obtained by the distributed DE are compared against those of the Grid Search as well as against those achieved by 13 well-known classification tools. The use of a distributed DE version turns out to be very effective in automatically obtaining DNN structures with higher classification accuracy with respect to Grid Search (72.95% versus 72.61%), and allows saving a high amount of time (three hours as opposed to 65 h and 40 min). Moreover, the proposed method outperforms in terms of higher accuracy all the other classifiers investigated, as it is evidenced also by statistical analysis. Numerically, the runner-up, i.e., JRip, achieves as its best value 72.01% and 71.50% on average over 25 runs, both values being lower than 72.95% and 72.74% obtained by our dDE.

Background and Objectives: Image segmentation represents one of the most challenging issues in medical image analysis to distinguish among different adjacent tissues in a body part. In this context, appropriate image pre-processing tools can improve the result accuracy achieved by computer-assisted segmentation methods. Taking into consideration images with a bimodal intensity distribution, image binarization can be used to classify the input pictorial data into two classes, given a threshold intensity value. Unfortunately, adaptive thresholding techniques for two-class segmentation work properly only for images characterized by bimodal histograms. We aim at overcoming these limitations and automatically determining a suitable optimal threshold for bimodal Magnetic Resonance (MR) images, by designing an intelligent image analysis framework tailored to effectively assist the physicians during their decision-making tasks.Methods: In this work, we present a novel evolutionary framework for image enhancement, automatic global thresholding and segmentation, which is here applied to different clinical scenarios involving bimodal MR image analysis: (i) uterine fibroid segmentation in MR guided Focused Ultrasound Surgery, and (ii) brain metastatic cancer segmentation in neuro-radiosurgery therapy. Our framework exploits MedGA as a pre-processing stage. MedGA is an image enhancement method based on Genetic Algorithms that improves the threshold selection, obtained by the efficient Iterative Optimal Threshold Selection algorithm, between the underlying sub-distributions in a nearly bimodal histogram.Results: The results achieved by the proposed evolutionary framework were quantitatively evaluated, showing that the use of MedGA as a pre-processing stage outperforms the conventional image enhancement methods (i.e., histogram equalization, bi-histogram equalization, Gamma transformation, and sigmoid transformation), in terms of both MR image enhancement and segmentation evaluation metrics.Conclusions: Thanks to this framework, MR image segmentation accuracy is considerably increased, allowing for measurement repeatability in clinical workflows. The proposed computational solution could be well-suited for other clinical contexts requiring MR image analysis and segmentation, aiming at providing useful insights for differential diagnosis and prognosis.

Many studies have shown that epicardial fat is associated with a higher risk of heart diseases. Accurate epicardial adipose tissue quantification is still an open research issue. Considering that manual approaches are generally user-dependent and time-consuming, computer-assisted tools can considerably improve the result repeatability as well as reduce the time required for performing an accurate segmentation. Unfortunately, fully automatic strategies might not always identify the Region of Interest (ROI) correctly. Moreover, they could require user interaction for handling unexpected events. This paper proposes a semi-automatic method for Epicardial Fat Volume (EFV) segmentation and quantification. Unlike supervised Machine Learning approaches, the method does not require any initial training or modeling phase to set up the system. As a further key novelty, the method also yields a subdivision into quartiles of the adipose tissue density. Quartile-based analysis conveys information about fat densities distribution, enabling an in-depth study towards a possible correlation between fat amounts, fat distribution, and heart diseases. Experimental tests were performed on 50 Calcium Score (CaSc) series and 95 Coronary Computed Tomography Angiography (CorCTA) series. Area-based and distance-based metrics were used to evaluate the segmentation accuracy, by obtaining Dice Similarity Coefficient (DSC) = 93.74% and Mean Absolute Distance (MAD) = 2.18 for CaSc, as well as DSC = 92.48% and MAD = 2.87 for CorCTA. Moreover, the Pearson and Spearman coefficients were computed for quantifying the correlation between the ground-truth EFV and the corresponding automated measurement, by obtaining 0.9591 and 0.9490 for CaSc, and 0.9513 and 0.9319 for CorCTA, respectively. In conclusion, the proposed EFV quantification and analysis method represents a clinically useable tool assisting the cardiologist to gain insights into a specific clinical scenario and leading towards personalized diagnosis and therapy.

Medical imaging systems often require the application of image enhancement techniques to help physicians in anomaly/abnormality detection and diagnosis, as well as to improve the quality of images that undergo automated image processing. In this work we introduce MedGA, a novel image enhancement method based on Genetic Algorithms that is able to improve the appearance and the visual quality of images characterized by a bimodal gray level intensity histogram, by strengthening their two underlying sub-distributions. MedGA can be exploited as a pre-processing step for the enhancement of images with a nearly bimodal histogram distribution, to improve the results achieved by downstream image processing techniques. As a case study, we use MedGA as a clinical expert system for contrast-enhanced Magnetic Resonance image analysis, considering Magnetic Resonance guided Focused Ultrasound Surgery for uterine fibroids. The performances of MedGA are quantitatively evaluated by means of various image enhancement metrics, and compared against the conventional state-of-the-art image enhancement techniques, namely, histogram equalization, bi-histogram equalization, encoding and decoding Gamma transformations, and sigmoid transformations. We show that MedGA considerably outperforms the other approaches in terms of signal and perceived image quality, while preserving the input mean brightness. MedGA may have a significant impact in real healthcare environments, representing an intelligent solution for Clinical Decision Support Systems in radiology practice for image enhancement, to visually assist physicians during their interactive decision-making tasks, as well as for the improvement of downstream automated processing pipelines in clinically useful measurements.

Prostate cancer is the most common malignant tumors in men but prostate Magnetic Resonance Imaging (MRI) analysis remains challenging. Besides whole prostate gland segmentation, the capability to differentiate between the blurry boundary of the Central Gland (CG) and Peripheral Zone (PZ) can lead to differential diagnosis, since the frequency and severity of tumors differ in these regions. To tackle the prostate zonal segmentation task, we propose a novel Convolutional Neural Network (CNN), called USE-Net, which incorporates Squeeze-and-Excitation (SE) blocks into U-Net, i.e., one of the most effective CNNs in biomedical image segmentation. Especially, the SE blocks are added after every Encoder (Enc USE-Net) or Encoder-Decoder block (Enc-Dec USE-Net). This study evaluates the generalization ability of CNN-based architectures on three T2-weighted MRI datasets, each one consisting of a different number of patients and heterogeneous image characteristics, collected by different institutions. The following mixed scheme is used for training/testing: (i) training on either each individual dataset or multiple prostate MRI datasets and (ii) testing on all three datasets with all possible training/testing combinations. USENet is compared against three state-of-the-art CNN-based architectures (i.e., U-Net, pix2pix, and Mixed-Scale Dense Network), along with a semi-automatic continuous max-flow model. The results show that training on the union of the datasets generally outperforms training on each dataset separately, allowing for both intra-/cross-dataset generalization. Enc USE-Net shows good overall generalization under any training condition, while Enc-Dec USE-Net remarkably outperforms the other methods when trained on all datasets. These findings reveal that the SE blocks' adaptive feature recalibration provides excellent cross-dataset generalization when testing is performed on samples of the datasets used during training. Therefore, we should consider multi-dataset training and SE blocks together as mutually indispensable methods to draw out each other's full potential. In conclusion, adaptive mechanisms (e.g., feature recalibration) may be a valuable solution in medical imaging applications involving multi-institutional settings. (C) 2019 Elsevier B.V. All rights reserved.

We present an unsupervised sclera segmentation method foreye color images. The proposed approach operates on a visible spectrumRGB eye image and does not require any prior knowledge such as eyelidor iris center coordinate detection. The eye color input image is enhancedby an adaptive histogram normalization to produce a gray level image inwhich the sclera is highlighted. A feature extraction process is involvedboth in the image binarization and in the computation of scores to assignto each connected components of the foreground. The binarization processis based on clustering and adaptive thresholding. Finally, the selectionof foreground components identifying the sclera is performed on theanalysis of the computed scores and of the positions between the foregroundcomponents. The proposed method was ranked 2nd in the ScleraSegmentation and Eye Recognition Benchmarking Competition (SSRBC2017), providing satisfactory performance in terms of precision.

Retinitis Pigmentosa is an eye disease that presents with aslow loss of vision and then evolves until blindness results. Theautomatic detection of the early signs of retinitis pigmentosa acts as agreat support to ophthalmologists in the diagnosis and monitoring of thedisease in order to slow down the degenerative process.A large body of literature is devoted to the analysis of RetinitisPigmentosa. However, all the existing approaches work on OpticalCoherence Tomography (OCT) data, while hardly any attempts have been madeworking on fundus images. Fundus image analysis is a suitable tool indaily practice for an early detection of retinal diseases and themonitoring of their progression. Moreover, the fundus camera represents alow-cost and easy-access diagnostic system, which can be employed inresource-limited regions and countries.The fundus images of a patient suffering from retinitis pigmentosa arecharacterized by an attenuation of the vessels, a waxy disc pallor andthe presence of pigment deposits. Considering that several methods havebeen proposed for the analysis of retinal vessels and the optic disk,this work focuses on the automatic segmentation of the pigment depositsin the fundus images. The image distortions are attenuated by applying alocal {\color{blue}pre-processing}. Next, a watershed transformation iscarried out to produce homogeneous regions. Working on regions ratherthan on pixels makes the method very robust to the high variability ofpigment deposits in terms of color and shape, so allowing the detectioneven of small pigment deposits. The regions undergo a feature extractionprocedure, so that a region classification process is performed by meansof an outlier detection analysis and a rule set. The experiments havebeen performed on a dataset of images of patients suffering fromretinitis pigmentosa. Although the images present a high variability interms of color and illumination, the method provides a good performancein terms of sensitivity, specificity, accuracy and the F-measure, whosevalues are 74.43, 98.44, 97.90, 59.04, respectively.

New technological advancements in automated microscopy gave rise to large volumes of data, which have made human-based analysis infeasible and have heightened the need of automatic systems for high-throughput microscopy applications. In particular in the field of fluorescence microscopy, automatic tools for image analysis make an essential contribution to increase the statistical power of the cell analysis process. This is a difficult task due to both diversification of staining patterns and local variability of the images. In order to cope with this challenges, this method divides the whole image in overlapping patches, which then undergo a gray level clustering. Thus, adaptive thresholding on clustered gray levels is applied to single patches to extract the foreground, while a merging process is implemented to recompose the foreground of overlapping patches into a final binary mask. The foreground represents the input of the cell counting stage. Since separating clustered cells is crucial for an accurate cell counting, centres of cells are detected by a two-stage process that combines the distance transform and curvature analysis. Labeling the detected centres, a partition of the image in single cells is obtained. The method has been tested on several publicly available image datasets with respect to both segmentation and cell counting. This package offers a powerful tool for automatic cells segmentation and counting. The foreground detection algorithm as well as the cell counting process is driven by a few number of parameters, whose purpose is detailed in the following. The image pre-processing is an optional step regulated by a boolean parameter, that is set to "true", when image correction is required and is "false", otherwise. The foreground detection is based on a sliding window, whose behavior is determined by two parameters that are the window size "n" and the sliding step "beta". The patch size "n" strongly depends on the resolution and homogeneity of the input image. Indeed, the more homogeneous are the objects into the image, the larger the value that can be assigned to "n". The parameter "beta" determines the degree of overlap of different patches. An appropriate value for "beta" can be selected according to the same heuristic adopted for "n". In the binarization process, a key role is played by the parameter "alpha", as it regulates the adapthive thresholding of the quantized patches. In particular, we notice that higher values of "alpha" must be set when pre-processing is applied to the image. This is motivated by the fact that pre-processing produces significant changes in the image contrast by spreading the original grey levels on a larger range of values. On the contrary, when pre-processing is omitted, the quantized patches are generally characterized by low contrast that induces smaller values for the parameter "alpha". The parameter "lambda" is also involved in the binarization process and regulates how much the thresholding of the current quantized patch is influenced by threshold values adopted for the preceding ones. The value of the parameter epsilon depends on the size of the smallest cell into the images. The only parameter involved in the counting process is "delta", which drives the incremental clustering of seeds produced by the distance transform. The value of this parameter strongly depends on the size of cells. Indeed, the smaller the cells, the lower the value of "delta" to be set, to avoid that different cells into a cluster will be merged.

The automatic segmentation of variations in fundus images is being increasingly developed to establish retinal health, and to diagnose and monitor retinal abnormality. Fundus images can be acquired by low-cost and easy access diagnostic systems also in resource limited countries. In this paper, we focus on the segmentation of pigment signs in retinal fundus images, which is an important step in the diagnosis and monitoring of Retinitis Pigmentosa. At present, most of the existing approaches adopted for this purpose are based on the analysis of Optical Coherence Tomography (OCT) data, with only a few algorithms working on fundus images. The contribution of this paper is twofold. First, we propose a supervised segmentation technique for pigment sign detection in fundus images, which exploits the ensemble classifiers for both bagged and boosted decision trees, namely Random Forests and AdaBoost.M1. The ensemble classifiers are trained on feature vectors encoding the information of a set of regions obtained by partitioning a pre-processed image, so that each region is labeled as either normal fundus or pigment sign. Secondly, we have collected a large dataset of retinal images, in which the pigment signs have been segmented manually and we are making this publicly available to the research community. The performance of the classifiers has been evaluated on the proposed dataset and can be considered as a baseline for comparison. Our results demonstrate the high effectiveness of a machine learning approach as a suitable tool for automated Retinitis Pigmentosa analysis.

The adoption of automatic systems to support the diagnosis of breast cancer from histology images analysis is rapidly becoming more widespread. Most of the works in literature focus principally on a two-class problem, namely benign and malignant tumors. However, the development of multi-classification approaches would also be greatly appreciated in order to support the determination of an ideal therapeutic schedule for the treatment of breast cancer. The multi-classification of histology images is particularly challenging due to the broad variability of appearance of the image, the great differences in the spatial arrangement of the histological structures, and the heterogeneity in the color distribution. In this work, a fine-tuning strategy of ResNet, a residual convolutional neural network, is presented to address the problem of multi-classification for breast cancer histology images in normal tissue, benign lesions, in situ carcinomas and invasive carcinomas.We have combined three configurations of ResNet, differing from each other in terms of the number of layers, by using a maximum probability rule to balance out their individual weaknesses during the testing. The proposed approach achieved a remarkable performance on the images provided for the Grand Challenge on Breast Cancer Histology Images (BACH), within the context of the International Conference ICIAR 2018.

We present a supervised method for vessel segmentation in retinal images. The segmentation issue has been addressed as a pixellevel binary classification task, where the image is divided into patches and the classification (vessel or non-vessel) is performed on the central pixel of the patch. The input image is then segmented by classifying all of its pixels. A Convolutional Neural Network (CNN) has been used for the classification task, and the network has been trained on a large number of samples, in order to obtain an adequate generalization ability. Since blood vessels are characterized by a linear structure, we have introduced a further layer into the classic CNN including directional filters. The method has been tested on the DRIVE dataset producing satisfactory results, and its performance has been compared to that of other supervised and unsupervised methods.

The wide availability of sensing devices in the medical domain causes the creation of large and very large data sets. Hence, tasks as the classification in such data sets becomes more and more difficult. Deep Neural Networks (DNNs) are very effective in classification, yet finding the best values for their hyper-parameters is a difficult and time-consuming task. This paper introduces an approach to decrease execution times to automatically find good hyper-parameter values for DNN through Evolutionary Algorithms when classification task is faced. This decrease is obtained through the combination of two mechanisms. The former is constituted by a distributed version for a Differential Evolution algorithm. The latter is based on a procedure aimed at reducing the size of the training set and relying on a decomposition into cubes of the space of the data set attributes. Experiments are carried out on a medical data set about Obstructive Sleep Anpnea. They show that sub-optimal DNN hyper-parameter values are obtained in a much lower time with respect to the case where this reduction is not effected, and that this does not come to the detriment of the accuracy in the classification over the test set items.

This paper describes our preliminary steps towards the deployment of a brand-new original feature for a telemedicine portal aimed at helping people suffering from diabetes. In fact, people with diabetes necessitate careful handling of their disease to stay healthy. As such a disease is correlated to a malfunction of the pancreas that produces very little or no insulin, a way to enhance the quality of life of these subjects is to implement an artificial pancreas able to inject an insulin bolus when needed. The goal of this paper is to extrapolate a regression model, capable of estimating the blood glucose (BG) through interstitial glucose (IG) measurements, that represents a possible revolutionizing step in constructing the fundamental element of such an artificial pancreas. In particular, a new evolutionary approach is illustrated to stem a mathematical relationship between BG and IG. To accomplish the task, an automatic evolutionary procedure is also devised to estimate the missing BG values within the investigated real-world database made up of both BG and IG measurements of people suffering from Type 1 diabetes. The discovered model is validated through a comparison with other models during the experimental phase on global and personalized data treatment. Moreover, investigation is performed about the accuracy of one single global relationship model for all the subjects involved in the study, as opposed to that obtained through a personalized model found for each of them. Once this research is clinically validated, the important feature of estimating BG will be added to a web portal for diabetic subjects for telemedicine purposes.

Deep Neural Networks are being more and more widely used to perform several tasks over highly-sized datasets, one of them being classification. Finding good configurations for Deep Neural Network structures is a very important problem in general, and particularly in the medical domain. Currently, either trial-and-error methodologies or sampling-based ones are considered. This paper describes some preliminary steps towards effectively facing this task. The first step consists in the use of Differential Evolution, a kind of an Evolutionary Algorithm. The second lies in using a parallelized version in order to reduce the turnaround time. The preliminary results obtained here show that this approach can be useful in easily obtaining structures that allow increases in the network accuracy with respect to those provided by humans.

An unsupervised method is introduced for retinal blood vessels segmentation. The direction map is built by assigning to each pixel a discrete direction out of twelve possible ones. Under- and over-segmented images are obtained by applying two different threshold values to the direction map. Almost all foreground pixels in the under-segmented image can be taken as vessel pixels. Missing vessel pixels in the under-segmented image are recovered by using the over-segmented image. The method has been tested on the DRIVE dataset producing satisfactory results, and its performance has been compared to that of other unsupervised methods.

A person verification algorithm involving retina segmentation, co-registration, feature extraction and matching.

Biometric systems based on retinal image analysis are of interest for highly accurate person verification. In this framework, we present a person verification method involving retina segmentation, co-registration, feature extraction and matching. All processes are based on the construction and use of a direction map and of an associated score map. An interesting feature of the method is the possibility to accomplish satisfactorily person verification also when the retinal image is not fully available at high quality, as it could be the case due to aging or retinal diseases of the subject. The proposed approach has been tested on the two publicly available datasets specifically created for evaluating biometric systems for retinal image based person verification, VARIA and RIDB. The approach has been compared with the other verification methods in the recent literature for which an evaluation on the above datasets is available. The experimental results confirm the goodness of the proposed approach and show that direction and score maps have a key role in person verification.

Color vision deficiency (CVD) is an extremely frequent vision impairment that compromises the ability to recognize colors. In order to improve color vision in a subject with CVD, we designed and developed a wearable improved vision system based on an augmented reality device. The system was validated in a clinical pilot study on 24 subjects with CVD (18 males and 6 females, aged 37.4 ± 14.2 years). The primary outcome was the improvement in the Ishihara Vision Test score with the correction proposed by our system. The Ishihara test score significantly improved (p=0.03) from 5.8 ± 3.0 without correction to 14.8 ± 5.0 with correction. Almost all patients showed an improvement in color vision, as shown by the increased test scores. Moreover, with our system, 12 subjects (50%) passed the vision color test as normal vision subjects. The development and preliminary validation of the proposed platform confirm that a wearable augmented-reality device could be an effective aid to improve color vision in subjects with CVD.

In the present work, a methodology for the analysis of subcellular morphology with chemical specificity for images from Stimulated Raman Scattering is suggested. In particular, a segmentation method based on a threshold algorithm and on a region growing process, to detect microstructures inside the cells, is proposed. Moreover, quantitative features for the segmented objects are extracted, in order to provide information about the possible morphological variations of microstructures in images acquired by means SRS technique.

Skin detection is an important process in many applications like hand gesture recognition, face detection and ego-vision systems. This paper presents a new skin detection method based on a dynamic generation of the skin cluster range in the YCbCr color space, by taking into account the lighting conditions. The method is based on the identification of skin color clusters in the YCb and YCr subspaces. The experimental results, carried out on two publicly available databases, show that the proposed method is robust against illumination changes and achieves satisfactory results in terms of both qualitative and quantitative performance evaluation parameters.

The cultural heritage could benefit significantly from the integration of wearable augmented reality (AR). This technology has the potential to guide the user and provide her with both in-depth information, without distracting her from the context, and a natural interaction, which can further allow her to explore and navigate her way through a huge amount of cultural information. The integration of touchless interaction and augmented reality is particularly challenging. On the technical side, the human-machine interface has to be reliable so as to guide users across the real world, which is composed of cluttered backgrounds and severe changes in illumination conditions. On the user experience side, the interface has to provide precise interaction tools while minimizing the perceived task difficulty. In this study, an interactive wearable AR system to augment the environment with cultural information is described. To confer robustness to the interface, a strategy that takes advantage of both depth and color data to find the most reliable information on each single frame is introduced. Moreover, the results of an ISO 9241-9 user study performed in both indoor and outdoor conditions are presented and discussed. The experimental results show that, by using both depth and color data, the interface can behave consistently in different indoor and outdoor scenarios. Furthermore, the results show that the presence of a virtual pointer in the augmented visualization significantly reduces the users error rate in selection tasks.

This paper presents a novel rule-based skin detection method that works in the YCbCr color space. The method is based on correlation rules that evaluate the combinations of chrominance values to identify the skin pixels in the YCb and YCr subspaces. The correlation rules depend on the shape and size of dynamically generated skin color clusters, which are computed on a statistical basis in the YCb and YCr subspaces for each single image, and represent the areas that include most of the candidate skin pixels. Comparisons with six well-known rule-based methods in literature carried out on four publicly available databases show that the proposed method outperforms the others in terms of quantitative performance evaluation parameters. Moreover, the qualitative analysis shows that the method achieves satisfactory results also in critical scenarios, including severe variations in illumination conditions.

In recent years finger vein authentication has gained an increasing attention, since it has shown the potential of providing high accuracy as well as robustness to spoofing attacks. In this paper we presented a new finger verification approach, which does not need precise segmentation of regions of interest (ROIs), as it exploits a co-registration process between two vessel structures. We tested the verification performance on the MMCBNU_6000 finger vein dataset, showing that this approach outperforms state of the art techniques in terms of Equal Error Rate.

The aim of this study is to investigate the correlation between, on the one hand, personal and life-style indicators and, on the other hand, the risk of falling. As indicators we consider here for each subject age, body mass index, and information about physical activity habits, while a subject's risk of falling is estimated by the Mini-BES test score. Three different groups of subjects are taken into account, namely healthy, suffering from metabolic diseases and suffering from cardiovascular diseases. Firstly, we aim at finding explicit linear correlations for any pair of parameters. Secondly, we wish to pay attention to whether or not these correlations change as the health state of the subjects does. The final goal is to move the first steps towards the design of a system composed by wearable sensors, a mobile device, and an app that would be able to help people in improving their life-style so as to decrease their falling risk.

Fall detection is very important for the health care especially for elderly people. The automatic discovery of falls in real time with the ability to differentiate them from normal daily activities is crucial. To achieve this aim, this paper proposes an approach based on getting data through a tag placed on the subject's chest, a windowing of the data, the automatic extraction through the DEREx tool of a set of IF-THEN rules able to classify windows as being part of fall or non-fall actions, and a final window composition to assess whether or not each global action was a fall. The approach is then tested on a real-world database containing a set of fall and non-fall actions, and is compared, in terms of classification over windows, against four state-of-the-art machine learning methods. Moreover, its results are also compared, in terms of accuracy in discrimination of the fall actions from the non-fall ones, against those obtained by the database builders through the use of another powerful machine learning algorithm. Numerical results are encouraging, and suggest that the proposed methodology could put solid ground for the design and the implementation of a real-time system for fall detection.

Subjects suffering from Type 1 diabetes mellitus need to constantly receive insulin injections. To improve their life quality, a desirable solution is represented by the implementation of an artificial pancreas. In this paper we move a preliminary step towards this goal. Namely, we work at the knowledge base for such a device. One of the main problems is to estimate the Blood Glucose (BG) values, starting from the easily available Interstitial Glucose (IG) ones, and this is the aim of our paper. To face this regression task we avail ourselves of Genetic Programming over a real-world database containing both BG and IG measurements for several subjects suffering from Type 1 diabetes, aiming at finding an explicit relationship between BG and IG values under the form of a mathematical expression. This latter could be the core of the knowledge base part of an artificial pancreas. Experimental comparisons against the state-of-the-art models evidence the quality of the proposed approach.

A novel skin detection method based on a dynamic generation of the skin cluster range in the YCbCr color space by taking into account the lighting conditions. The method is based both on the identification of skin color clusters in the YCb and YCr subspaces and on the definition of correlation rules between the skin color clusters; it is efficient in terms of computational effort and is robust against severe variation in illumination conditions.

Understanding the mechanisms of invasion of cancer cells into surrounding tissues is of primary importance for limiting tumor progression. The degradation of the extracellular matrix (ECM) and the consequent invasion of the surrounding tissue by tumor cells represent the first stage in the development and dissemination of metastasis. The quantification of such a degradation is thus an important parameter to evaluate the metastatic potential of tumor cells. Assessment of degradation is usually performed in in vitro assays, in which tumor cells are cultured on a gelatin (or other matrix)-coated dishes and the degraded gelatin areas under the tumor cells are visualized and quantified by fluorescent labelling. In this paper, we present an automatic method to quantify the ECM degradation through the feature analysis of the digital images, obtained from the in vitro assays and showing the tumor cells and the degraded gelatin areas. Differently from the existing methods of image analysis supporting biologists, our method does not require any interaction with the user providing quickly corrected and unbiased measures. Comparative results with a method frequently used by biologists, has been performed.

This paper presents a new approach for skin detection in colour images. The method is based on the building of a dynamic clustering in the YCbCr colour space, taking into account the illumination conditions of the examined image. The results of a comparative evaluation on a publicly available database, show that the proposed method outperforms well known rule based static methods, both in qualitative and quantitative terms.

RGB-D Fingertip Detection Ego frames. The dataset has been collected from 3 subjects. For each subject, the acquisition was performed in uncontrolled environments and under different lighting conditions measured in lumens. In total the dataset contains 10 videos, each acquired at 30 fps and lasting 10 seconds, for a total of 300 frames for each video.

This paper presents the unsupervised retinal vessels segmentation method, SEVERE (SEgmenting VEssels in REtina images), which is based on the direction map of retina scan images assigning each pixel one out of twelve discrete directions. SEVERE works on the green channel of RGB retina scan images. It does not require any pre-processing phase and all the computations are done exclusively on the direction map. SEVERE has been checked on publicly available datasets producing qualitatively satisfactory results and outperforming other existing methods in terms of quantitative performance evaluation parameters, such as accuracy and sensitivity. (C) 2015 Elsevier B.V. All rights reserved.

A Watershed transform based Iris REcognition system (WIRE) for noisy images acquired in visible wavelength is presented. Key points of the system are: the color/illumination correction pre-processing step, which is crucial for darkly pigmented irises whose albedo would be dominated by corneal specular reflections; the criteria used for the binarization of the watershed transform, leading to a preliminary segmentation which is refined by taking into account the watershed regions at least partially included in the best iris fitting circle; the introduction of a new cost function to score the circles detected as potentially delimiting limbus and pupil. The advantage offered by the high precision of WIRE in iris segmentation has a positive impact as regards the iris code, which results to be more accurately computed, so that also the performance of iris recognition is improved. To assess the performance of WIRE and to compare it with the performance of other available methods, two well known databases have been used, specifically UBIRIS version 1 session 2 and the subset of UBIRIS version 2 that has been used as training set for the international challenge NICE II.

The aim of this study is to identify an explicit relationship between life-style and the risk of falling under the form of a mathematical model. Starting from some personal and behavioral information as, e.g., weight, height, age, data about physical activity habits, and concern about falling, the model would easily estimate the score of the Mini-Balance Evaluation Systems (Mini-BES) test. This would make fall risk assessment less invasive, because subjects would not need to undergo the classical Mini-BES test, rather they could estimate it at home by answering some questionnaires. The mathematical model obtained in this study has been tested over a subset of unseen subjects and the results show an average error of +-2.74.

Continuous blood pressure (BP) measurement is an important issue in the medical field. The hypothesis of existence of a nonlinear relationship between plethysmography (PPG) and BP values has been investigated in this paper. If this hypothesis is true, then it is possible to indirectly measure patient's BP in a non-invasive way through the application of a wearable wireless PPG sensor to patient's finger and through the use of the results of a regression analysis aimed at linking PPG and BP values. To find the relationship between these two biomedical characteristics we have used here Genetic Programming (GP), because in a regression task it can evolve in an automatic way the structure of the most suitable explicit mathematical model. An analysis of the related scientific literature shows that this is the first attempt to mathematically relate PPG and BP values through GP. In this paper some preliminary experiments on the use of GP in facing this regression task have been carried out. As a result, for both systolic and diastolic BP values explicit mathematical models providing nonlinear relationship between PPG and BP values have been achieved, involving an approximation error of around 2 mmHg in both cases. A prototypal mobile-based system has been realized which is able to continuously estimate in real time the two BP values for any given patient by using only a plethysmography signal and the obtained mathematical models.

This paper investigates the hypothesis that a nonlinear relationship exists between photoplethysmography (PPG) and blood pressure (BP) values. Trueness of this hypothesis would imply that, instead of measuring a patient's BP in an invasive way, this could be indirectly measured by applying a wearable PPG sensor and by using the results of a regression analysis linking PPG and BP. Genetic Programming (GP) is well suited to find the relationship between PPG and BP, because it automatically evolves the structure of the most suitable explicit mathematical model for a regression task. In this paper, for the first time, some preliminary experiments on the use of GP to explicitly relate PPG and BP values have been performed. For both systolic and diastolic BP values, explicit nonlinear mathematical models have been achieved, involving an approximation error of less than 3 mmHg in both cases.

Wearable augmented reality (AR) systems have the potential to significantly lower the barriers to accessing information, while leaving the focus of the user's attention on the real world. To reveal their true potential, the human-machine interface is crucial. A touchless point-and-click interface for wearable AR systems may be suitable for use in many realworld applications, but it demands fingertip detection techniques robust enough to cope with cluttered backgrounds and varying illumination conditions. In this paper we propose an approach that, by automatically choosing between color and depth features, allows to detect the hand and then the user's fingertip both in indoor and outdoor scenarios, with or without adequate illumination.

La fruizione del patrimonio culturale, tangibile e intangibile, è oggi in forte evoluzione. Il visitatore può non solo guardare le opere, ma interagire, richiedere informazioni aggiuntive su ciò che vede, inquadrare l'opera nel contesto socio-culturale. Le tecnologie di realtà aumentata e virtuale stanno diventando sempre più un valido strumento per rispondere a queste esigenze. Tuttavia, affinché tali tecnologie divengano un elemento efficace per la fruizione del patrimonio culturale, devono poter essere semplici da usare, non ingombranti e poter supportare il visitatore, fornendogli informazioni in qualsiasi luogo, outdoor (e.g., siti turistici, piazze), o indoor (e.g., musei, chiese). A valle di una panoramica delle nuove tecnologie e delle loro potenzialità, in questo articolo viene presentato un sistema prototipale che, tramite un dispositivo indossabile di realtà aumentata adatto all'utilizzo sia indoor che outdoor, permette di interagire mediante comandi gestuali con le informazioni proiettate nel campo visivo dell'utente.

The paper deals with target selection techniques for wearable augmented reality systems. In particular, we focus on the three techniques most commonly used in distant freehand pointing and clicking on large displays: wait to click, air tap and thumb trigger. The paper details the design of the techniques for a touchless augmented reality interface and provides the results of a preliminary usability evaluation carried out in out-of-lab settings.

The recent availability of low cost wearable augmented reality (WAR) technologies is leveraging the design of applications in the cultural heritage domain in order to support users in their emotional journey among the cultural artefacts and monuments of a city. In this paper, we describe a user study evaluating the usability of a wearable augmented reality touchless interface for the enjoyment of the cultural heritage in outdoor environments. The usability evaluation has been carried out in out-of-lab settings with inexperienced users, during a three day exhibition in the city of Naples. The presented results are related to the ease of use and learning of the system, and to the user's satisfaction in the enjoyment of the system.

Communications with a central iris database system using common wireless technologies, such as tablets and smartphones, and iris acquisition out of the field are important functionalities and capabilities of a mobile iris identification device. However, when images are acquired by means of mobile devices under uncontrolled acquisition conditions, noisy images are produced and the effectiveness of the iris recognition system is significantly conditioned. This paper proposes a technique based on watershed transform for iris detection in noisy images captured by mobile devices. The method exploits the information related to limbus to segment the periocular region and merges its score with the iris' one to achieve greater accuracy in the recognition phase.

We present an effective unsupervised segmentation method that is based only on the use of the direction map built in correspondence of the retinal image by assigning each pixel one out of twelve discrete directions. The segmentation method works on the green channel of RGB retina images and does not require any pre-processing phase. The method has been checked on the DRIVE dataset producing satisfactory results both qualitatively and quantitatively.

Automatic fall detection is a major issue in the health care of elderly people. In this task the ability to discriminate in real time between falls and normal daily activities is crucial. Several methods already exist to perform this task, but approaches able to provide explicit formalized knowledge and high classification accuracy have not yet been developed and would be highly desirable. To achieve this aim, this paper proposes an innovative and complete approach to fall detection based both on the automatic extraction of knowledge expressed as a set of IF-THEN rules from a database of fall recordings, and on its use in a mobile health monitoring system. Whenever a fall is detected by this latter, the system can take immediate actions, e.g. alerting medical personnel. Our method can easily overcome the limitations of other approaches to fall detection. In fact, thanks to the knowledge gathering, it overcomes both the difficulty faced by a human being dealing with many parameters and trying to find out which are the most suitable, and also the need to apply a laborious trial-and-error procedure to find the values of the related thresholds. In addition, in our approach the extracted knowledge is processed in real time by a reasoner embedded in a mobile device, without any need for connection to a remote server. This proposed approach has been compared against four other classifiers on a database of falls simulated by volunteers, and its discrimination ability has been shown to be higher with an average accuracy of 91.88%. We have also carried out a very preliminary experimental phase. The best set of rules found by using the previous database has allowed us to achieve satisfactory performance in these experiments as well. Namely, on these real-world falls the obtained results in terms of accuracy, sensitivity, and specificity are of about 92%, 86%, and 96%, respectively. (C) 2015 Elsevier B.V. All rights reserved.

Obstructive Sleep Apnea (OSA) is a breathing disorder that takes place during sleep, and has both short- as well as long- term consequences on patient's health. Real-time monitoring for a patient can be carried out by making use of ElectroCardioGraphy (ECG) recordings. This paper introduces a methodology to forecast OSA events in the minutes following the current time instant. This is accomplished by using a tool based on Differential Evolution that is able to automatically extract offline knowledge about the monitored patient as a form of a set of IF-THEN rules. These rules connect the values of some ECG-related parameters recorded in the last minutes the occurrence of an apnea episode in the following minute. This approach has been tested on a literature database with 35 OSA patients. A comparison against six well-known classifiers has been performed.

This paper relies on the hypothesis of the existence of a nonlinear relationship between Electrocardiography (ECG) and Heart Related Variability (HRV) parameters, plethysmography (PPG), and blood pressure (BP) values. This hypothesis implies that, rather than continuously measuring the patient's BP, both their systolic and diastolic BP values can be indirectly measured as follows: a wearable wireless PPG sensor is applied to a patient's finger, an ECG sensor to their chest, HRV parameter values are computed, and regression is performed on the achieved values of these parameters. Genetic Programming (GP) is a Computational Intelligence paradigm that can at the same time automatically evolve the structure of a mathematical model and select from among a wide parameter set the most important parameters contained in the model. Consequently, it can carry out very well the task of regression. The scientific literature of this field reveals that nobody has ever used GP aiming at relating parameters derived from HRV analysis and PPG to BP values. Therefore, in this paper we have carried out preliminary experiments on the use of GP in facing this regression task. GP has been able to find a mathematical model expressing a nonlinear relationship between heart activity, and thus ECG and HRV parameters, PPG and BP values. The experimental results reveal that the approximation error involved by the use of this method is lower than 2 mmHg for both systolic and diastolic BP values.

In this paper an iris detection scheme for noisy images acquired by means of mobile devices is presented. Iris segmentation is accomplished by exploiting the use of the watershed transform with the purpose of identifying the iris boundary as much precisely as possible. After a pre-processing step aimed at color/illumination correction, the watershed transform is computed and suitably binarized. Circle fitting is then accomplished to identify the limbus boundary by using curvature approximation and a cost function for circle scoring. The watershed transform is furthermore employed to distinguish, in the zone delimited by the best fitting circle, the regions actually belonging to the iris from those belonging to eyelids and sclera. Finally, pupil detection is accomplished by means of circle fitting and by using a voting function based on homogeneity and separability criteria. The suggested iris detection scheme has a positive impact on an the accuracy in computing the iris code, which has in turn a positive impact on the performance of iris recognition.

In this paper we present a method to segment retinal vessels based on the use of the watershed transform and of the contrast map and the directional map built for the retina image. Pixels belonging to a single region of the watershed transform are assigned a unique gray-level and a unique direction. Each region of the watershed transform is also assigned a unique contrast value, computed as the maximum difference in gray-level with respect to its adjacent regions. Regions with positive value in the contrast map are interpreted as non vessel regions, while for the remaining regions directional information is used to identify the retinal vessels. The algorithm has been tested on the DRIVE database producing in general satisfactory results.

Remote monitoring of patients' vital parameters and ensuring mobility of both patient and doctor can greatly profit from real-time tele-monitoring technology. Here a description is given of a multi-purpose and multi-parametric tele-monitoring system. It can take advantage of the extraction, carried out offline and automatically on a desktop, of knowledge from databases containing measurements of patient's parameters. This knowledge is represented under the form of a set of IF...THEN rules that are provided to a rule-based mobile Decision Support System embedded in the system here presented. Then, wearable sensors collect in real time patient's vital parameters that are sent to a mobile device, where they are processed in real time by an app. If, as a consequence of the measured parameters, one of the above rules is activated, an alarm is automatically generated by the system for a well-timed medical intervention. Moreover all the monitored parameters are stored in EDF files for possible further analysis. This paper presents two practical applications of the system to two significant healthcare issues, i.e. apnea monitoring and fall detection. For these use cases, comparison with other well-known classifiers is carried out to evaluate the quality of the extracted knowledge.

Real-time Obstructive Sleep Apnea (OSA) episode detection and monitoring are important for society in terms of an improvement in the health of the general population and of a reduction in mortality and healthcare costs. Currently, to diagnose OSA patients undergo PolySomnoGraphy (PSG), a complicated and invasive test to be performed in a specialized center involving many sensors and wires. Accordingly, each patient is required to stay in the same position throughout the duration of one night, thus restricting their movements. This paper proposes an easy, cheap, and portable approach for the monitoring of patients with OSA, which collects single-channel ElectroCardioGram (ECG) data only. It is easy to perform from the patient's point of view because only one wearable sensor is required, so the patient is not restricted to keeping the same position all night long, and the detection and monitoring can be carried out in any place through the use of a mobile device. Our approach is based on the automatic extraction, from a database containing information about the monitored patient, of explicit knowledge in the form of a set of IF...THEN rules containing typical parameters derived from Heart Rate Variability (HRV) analysis. The extraction is carried out off-line by means of a Differential Evolution algorithm. This set of rules can then be exploited in the real-time mobile monitoring system developed at our Laboratory: the ECG data is gathered by a wearable sensor and sent to a mobile device, where it is processed in real time. Subsequently, HRV-related parameters are computed from this data, and, if their values activate some of the rules describing the occurrence of OSA, an alarm is automatically produced. This approach has been tested on a well-known literature database of OSA patients. The numerical results show its effectiveness in terms of accuracy, sensitivity, and specificity, and the achieved sets of rules evidence the user-friendliness of the approach. Furthermore, the method is compared against other well known classifiers, and its discrimination ability is shown to be higher.

Pruning is an important step in a skeletonization process and a number of pruning criteria have been suggested in the literature. However, the modality to be followed when checking the pruning criterion is not generally described in detail. In our opinion, two main pruning modalities can be envisaged and in this paper we discuss their impact on the performance of pruning. Moreover, we introduce a third modality, which we regard as able to provide a more satisfactory pruning performance.

Iris biometric systems are of interest for security applications. In this respect, iris segmentation has a key role, as it must be fast and accurate. In this paper, we present a new watershed based approach for iris segmentation in color images. The watershed transform is used in two distinct phases of iris segmentation: it is first used to obtain a preliminary segmentation, which constitutes the input to a circle fitting procedure; then, it is used together with the portion of the input image resulting after circle fitting to identify more precisely the pixels actually belonging to the iris. The experimental results show that the suggested approach is effective with respect to both location accuracy and computational complexity.

Recently, the research interest on biometric systems and applications has significantly grown up, aiming to bring the benefits of biometrics to the broader range of users. As signal processing and feature extraction play a very important role for biometric applications, they can be thought as a particular subset of pattern recognition techniques. The most of the iris biometric systems have been designed for security applications and work on near-infrared images. NIR images are not affected by illumination changes in visible light making systems working both in darker and lighter conditions. The reverse of the medal is a very short distance allowed between the acquisition camera and the user, further than a strictly controlled pose of the eye. For those reasons, the viability of NIR image based systems in commercial applications is quite limited. Several efforts have been devoted to designing new iris biometric approaches on color images acquired in visible wavelength light (VW). However, illumination changes significantly affect the iris pattern as well as the periocular region making both segmentation and feature extraction harder than in NIR. In the specific case of iris biometrics, segmentation represents a crucial aspect, as it must be fast as well as accurate. To this aim, a new watershed based approach for iris segmentation in color images is presented in this paper. The watershed transform is exploited to binarize an image of the eye, while circle fitting together with a ranking approach is applied to first approximate the iris boundary with a circle. The experimental results demonstrate this approach to be effective with respect to location accuracy.

This paper presents a method for automatically pair-wise registering range images. Registration is effected adding chaos to a Differential Evolution technique and by applying the Grid Closest Point algorithm to find the best possible transformation of the second image causing 3D reconstruction of the original object. Experimental results show the capability of the method in picking up efficient transformations of images with respect to the classical Differential Evolution. The proposed method offers a good solution to build complete 3D models of objects from 3D scan datasets.

In this paper, a new approach based on Differential Evolution (DE) for the automatic classification of items in medical databases is proposed. Based on it, a tool called DEREx is presented, which automatically extracts explicit knowledge from the database under the form of IF-THEN rules containing AND-connected clauses on the database variables. Each DE individual codes for a set of rules. For each class more than one rule can be contained in the individual, and these rules can be seen as logically connected in OR. Furthermore, all the classifying rules for all the classes are found all at once in one step. DEREx is thought as a useful support to decision making whenever explanations on why an item is assigned to a given class should be provided, as it is the case for diagnosis in the medical domain. The major contribution of this paper is that DEREx is the first classification tool in literature that is based on DE and automatically extracts sets of IF-THEN rules without the intervention of any other mechanism. In fact, all other classification tools based on DE existing in literature either simply find centroids for the classes rather than extracting rules, or are hybrid systems in which DE simply optimizes some parameters whereas the classification capabilities are provided by other mechanisms. For the experiments eight databases from the medical domain have been considered. First, among ten classical DE variants, the most effective of them in terms of highest classification accuracy in a ten-fold cross-validation has been found. Secondly, the tool has been compared over the same eight databases against a set of fifteen classifiers widely used in literature. The results have proven the effectiveness of the proposed approach, since DEREx turns out to be the best performing tool in terms of highest classification accuracy. Also statistical analysis has confirmed that DEREx is the best classifier. When compared to the other rule-based classification tools here used, DEREx needs the lowest average number of rules to face a problem, and the average number of clauses per rule is not very high. In conclusion, the tool here presented is preferable to the other classifiers because it shows good classification accuracy, automatically extracts knowledge, and provides users with it under an easily comprehensible form.

An intelligent system for supporting medical diagnosis is presented in this paper. The system automatically extracts knowledge from databases as sets of IF-THEN rules. The approach chosen to fulfi l this task is based on the differential evolution (DE) algorithm and its implementation results in a tool called DEREx. This tool is aimed at supporting clinicians in their decision making in the diagnostic process, by providing them with clear explanations on the reasons why each item is assigned to a given class. Performance of the tool has been evaluated over seven medical databases and compared against that of fi fteen well-known classification tools. Numerical results in terms of classifi cation accuracy and their statistical analysis, have evidenced the effectiveness of the proposed approach, so DEREx is preferable because of its added value, i.e. the knowledge extracted automatically and provided to users in an easily comprehensible form.

Performing detection and real-time monitoring of Obstructive Sleep Apnea (OSA) is a significant healthcare task. An easy, cheap, and mobile approach to monitor patients with OSA is proposed here. It gathers data from a patient by a single-channel ECG, and offline automatically extracts knowledge about that patient as a set of IF...THEN rules containing Heart Rate Variability (HRV) parameters. These rules are then used in the real-time mobile monitoring system: ECG data is collected by a wearable sensor, sent to a mobile device, and processed online to compute HRV-related parameter values. If a rule is activated by those values, the system produces an alarm. A literature database of OSA patients has been used to test the approach.

In this paper, we present a system for the generation of cultural itineraries that exploits conversational agents to implicitly build formal user profiles. The key idea is that the preferences for user profiling are not obtained in a direct way, but acquired during a natural language conversation of the tourists with the system. When the user profile is ready, it becomes the input for the generation of the customized cultural itinerary. The proposed system, called DiEM System, is designed for dialogues in the domain of cultural heritage, but its flexible architecture allows to customize the dialogues in different application domains (cinema, finance, medicine, etc.).

A fully automatic scaling up algorithm is presented in the framework of interpolation methods. For any integer zooming factor n, the algorithm generates a magnified version of an input color image in one scan of the image. The computational complexity of the algorithm is O(N), where N is the size of the input image. The visual aspect of the magnified images is generally appealing also when considering large zooming factors. Peak Signal to Noise Ratio and Structural SIMilarity are used to evaluate the performance of the algorithm and to compare it with other scaling up algorithms.

During the period 2010-2012, both theoretical and applicative aspects of the design and the implementation of discrete methods have been addressed. In particular, new tools to process and analyze 2D and 3D digital images have been developed. The main issues faced during the biennium deal with skeleton computation, object decomposition, image segmentation and image processing.

This paper presents an advanced ubiquitous system for home health monitoring. The main goal of the research presented in this paper is to develop a user-friendly and context-aware system that uses a rule-based Decision Support System to elaborate the data captured by the sensors. The paper also describes a case study where important benefits for patients have been revealed thanks to the use of the proposed home health monitoring system.

We present an interpolation algorithm for adaptive color image zooming. The algorithm produces the magnified image in one scan of the input image, and is fully automatic since does not involve any a priori fixed threshold. Given any integer zooming factor n, each pixel of the input image generates an nn block of pixels in the zoomed image. For the currently visited pixel of the input image, the pixels of its associated block are first assigned tentative values, which are then adaptively updated before building the nextblock. The method is suggested for RGB images, but can equally be employed in other color spaces. Peak signal to noise ratio (PSNR) and Structural SIMilarity (SSIM) are used to evaluate the performance of the algorithm.

A new segmentation method is presented. The watershed transformation is initially computed starting from all seeds detected as regional minima in the gradient image and a digging cost is associated to each pair of adjacent regions. Digging is performed for each pair of adjacent regions for which the cost is under a threshold, whose value is computed automatically, so originating a reduced set of seeds. Watershed transformation and digging are repeatedly applied, until no more seeds are filtered out. Then, region merging is accomplished, based on the size of adjacent regions.

Assisted Living provides a long-term care option that combines supportive systems andservices for monitoring and assessing the health status with activities of daily living andhealth care. Daily monitoring of the health status in subjects characterized by chronic and/ordegenerative conditions is not possible in all those cases where the disease progressionhas to be evaluated only by a direct interaction between the patients and the healthcarestructures on a regular basis, over time and for life. In this respect, this work proposes anevolutionary-fuzzy decision support system (DSS) for assessing the health status of subjectsaffected by multiple sclerosis (MS) during the disease progression over time. Such a DSS hasbeen defined and implemented exploiting a novel approach devised to facilitate the designof fuzzy DSSs for medical problems. The approach is aimed at: (i) introducing a set of designcriteria to encode the medical knowledge elicited from clinical experts in terms of linguisticvariables, linguistic values and fuzzy rules with the final aim of granting the interpretabil-ity; (ii) defining a fuzzy inference technique to best fit the structure of medical knowledgeand the peculiarities of the medical inference; (iii) defining an evolutionary technique totune the formalized knowledge by optimizing the shapes of the membership functions foreach linguistic variable involved in the rules. An experimental session has been carried outfor evaluating, first of all, the approach on five medical databases commonly diffused inliterature and for comparing it with other systems. After that, the evolutionary-fuzzy DSSfor assessing MS patient's health status has been quantitatively evaluated on 120 patientsaffected by MS and compared with other approaches. The achieved results have shown thatour approach is very effective on the five databases, since it provides, on average, the secondhighest accuracy when compared to eight tools. Furthermore, as far as the classification ofmultiple sclerosis lesions is considered, the proposed system has turned out to outperformnine popular tools.

The activity of the GIRPR research unit at the Institute of Cybernetics "E. Caianiello" - CNR is concerned with theoretical and applicative aspects of the design and the implementation of discrete methods and tools to analyze digital 2D and 3D gray-level and color images, at single or multiple scale. The main issues faced in the period 2008-2010 deal with medial representation of 3D objects, 2D gray-level image segmentation, 2D color image zooming and color quantization. Applications in the field of security have also been dealt with.

mHealth systems are becoming very attractive for the home care monitoring and, in particular, for the monitoring of patients with heart failure. Knowledge-based technologies can be profitably used to design advanced software system able to provide efficient and dependable service to patients and physicians. In this paper we present a Rule-based Decision Support System for mHealth environments; the designed intelligent system is devised to the detection and signaling of abnormal or emergency situations by using contextual information, i.e. by correlating data coming from a wearable electrocardiography (ECG) device with information regarding patient's posture and his/her physical activities. The whole system has been developed in Java.

Image segmentation methods involve a number of parameters whose values have to be tuned depending on image domain. In this communication, a watershedbased segmentation algorithm is considered and Case-Based-Reasoning is used for the automatic selection of the values that, assigned to the parameters, produce a satisfactory segmentation. In this way, the segmentation algorithm can be applied to a wider image domain.

In this paper, we present a new method for reconnecting broken ridges in fingerprint images. The method is based on the use of a discrete directional mask and on the standard deviation of the gray-levels to determine ridge direction. The obtained direction map is smoothed by counting the occurrences of the directions in a sufficiently large window. The fingerprint image is, then, binarized and thinned. Linking paths to connect broken ridges are generated by using a morphological transformation to guide the process.

In this chapter, a parallel software system based on differential evolution for the registration of images is designed, implemented and tested on a set of 2-D images in two different fields, i.e. remote sensing and medicine. Two different problems, i.e. mosaicking and changes in time, are faced in the former application field. Registration is carried out by finding the most suitable affine transformation in terms of maximization of the mutual information between the first image and the transformation of the second one, without any need for setting control points. A coarse-grained distributed version is implemented on a cluster of personal computers.

This chapter introduces a novel image-segmentation scheme based on case-based reasoning. Image segmentation is aimed at dividing an image into a number of different regions in such a way that each region is homogeneous with respect to a given property, but the union of any two adjacent regions is not. To reach this goal, a number of different approaches have been suggested in the literature, among which we consider here watershed-based segmentation. The basic idea of this segmentation scheme is to identify in the gray-level image a suitable set of seeds from which to perform a growing process. The growing process groups to each seed all pixels that are closer to that seed more than to any other seed, provided that a certain homogeneity condition is satisfied. Since any segmentation method includessome parameters, whose values depend on the image characteristics, CBR can be profitably used to improve the performance of the adopted segmentation method and to ensure that good segmentation results are obtained even if the segmentation method is applied to images with different characteristics. In practice, CBR will select from a case-base the cases having image characteristics similar to those of the current input image, and will apply to the current image the segmentation parameters associated to the most similar case. Image characteristics will be computed in terms of mean features on the whole image, and a proper similarity measure will be used to select in the case-base the most similar case.

In this paper, we present a new segmentation algorithm, based on iterated thresholding and on morphological features. A first thresholding, based on the histogram of the image, is done to partition the image into three sets including respectively pixels belonging to foreground, pixels belonging to background, and unassigned pixels. Thresholding of components of unassigned pixels is then iteratively done, based on the histogram of the components. Components of unassigned pixels, possibly still present at the end of iterated thresholding, are assigned to foreground or background by taking into account area, minimum grey-level and spatial relationship with the adjacent sets.

A new segmentation algorithm is suggested, which is based on iterated thresholding and on morphological features. The histogram of the grey-level image is used to identify two initial global thresholds ?1 and ?2, used to assign to the foreground and to the background respectively, pixels with grey-level below ?1 and above ?2. Local thresholding is then accomplished for each component of pixels that have not been assigned to any of the two sets by the global thresholding process. For each component, a new pair of thresholds is detected on the relative histogram. Local thresholding is applied to the components of undecided pixels as far as the relative histogram presents valleys and peaks. Then, to assign to the foreground or to the background the still undecided sets of pixels, morphological features are used. The suggested segmentation method works well for images, like many biological images, where the foreground is perceived as locally darker (or locally lighter) than the background, consistently through the whole image, and performs better than segmentation based on simple global thresholding.

This special issue constitutes a collection of extended versions of selected papers presented at the Second International Symposium on "Brain, Vision and Artificial Intelligence" (BVAI 2007), held in Xaples, Italy, on 10-12 October, 2007. The collected papers deal with novel techniques in Artificial Vision (AV) and Artificial Intelligence (AI), and with models inspù·ed by natural visìon and brain for AV and AI problems. The following are the two main thematic areas.

Automatic recognition of hand gestures is a crucial step in facing human-computer interaction. Differential Evolution is used to perform automatic classification of hand gestures in a thirteen-class database. Performance of the resulting best individual is computed in terms of error rate on the testing set, and is compared against those of other ten classification techniques well known in literature. Results show the effectiveness and the efficiency of the approach in solving the classification task. Furthermore, the implemented tool allows to extract the most significant parameters for differentiating the collected gestures.

Target behaviours can be achieved by finding suitable parameters for Continuous Time Recurrent Neural Networks (CTRNNs) used as agent control systems. Differential Evolution (DE) has been deployed to search parameter space of CTRNNs and overcome granularity, boundedness and blocking limitations. In this paper we provide initial support for DE in the context of two sample learning problems.

A software system grounded on Differential Evolution to automatically register multiview and multitemporal images is designed, implemented and tested through a set of 2D satellite images on two problems, i.e. mosaicking and changes in time. Registration is effected by looking for the best affine transformation in terms of maximization of the mutual information between the first image and the transformation of the second one, and no control points are needed in this approach. This method is compared against five widely available tools, and its effectiveness is shown.

In this paper we build a shape preserving resolution pyramid and use it in the framework of image segmentation via watershed transformation. Our method is based on the assumption that the most significant image components perceived at high resolution will also be perceived at lower resolution. Thus, we detect the seeds for the watershed transformation at a low resolution, and use them to distinguish significant and non-significant seeds at any selected higher resolution. In this way, the watershed partition obtained at the selected pyramid level will include only the most significant components, and over-segmentation will be considerably reduced. Segmentations of the image at different scales will be available. Moreover, since the seeds can be detected at different pyramid levels, alternative segmentations of the image at a given resolution can be obtained, each characterized by a different level of detail.

Gray-level image segmentation is the first task for any image analysis process, and is necessary to distinguish the objects of interest from the background. Segmentation is a complex task, especially when the gray-level distribution along the image is such that sets of pixels characterized by a given gray-level are interpreted by a human observer as belonging to the foreground in certain parts of the image, and to the background in other parts, depending on the local context. It very seldom happens that the background is characterized by an almost uniform gray-level. Thus, in the majority of cases, segmentation cannot be achieved by simply thresholding the image, i.e., by assigning all pixels with gray-level lower than a given threshold to the background and all remaining pixels to the foreground. One of the most often adopted segmentation techniques is based on a preliminary partition of the input gray-level image into regions, homogeneous with respect to a given property, to successively classify the obtained regions in two classes (foreground and background). In this paper, we follow this approach and present a powerful method to discriminate regions in a partition of a gray-level image obtained by using the watershed transformation. The basic idea underlying the classification is that for a wide class of graylevel images, e.g., a number of biological images, the boundary between the foreground and the background is perceived where locally maximal changes in gray-level occur through the image. Our classification procedure works well even starting from a standard watershed partition, i.e., without resorting to seed selection and region growing. However, we will also briefly discuss new criteria to be used when applying digging and flooding techniques in the framework of watershed transformation, so as to produce a less fragmented partition of the image. By using the so obtained partition of the gray-level image, the successive classification is facilitated and the quality of the obtained results is improved. Some hints regarding the use of multi-scale image representation to reduce the computational load will also be introduced.

This paper proposes a novel grey-level image segmentation scheme employing case-based reasoning. Segmentation is accomplished by using the watershed transformation, which provides a partition of the image into regions whose contours closely fit those perceived by human users. Case-based reasoning is used to select the segmentation parameters involved in the segmentation algorithm by taking into account the features characterizing the current image. Preliminarily, a number of images are analyzed and the parameters producing the best segmentation for each image, found empirically, are recorded. These images are grouped to form relevant cases, where each case includes all images having similar image features, under the assumption that the same segmentation parameters will produce similarly good segmentation results for all images in the case.

This paper deals with the design and implementation of a software system based on Differential Evolution for the registration of images, and in its testing by means of a set of bidimensional remotely sensed images on two problems, i.e. mosaicking and changes in time. Registration is carried out by finding the most suitable affine transformation in terms of maximization of the mutual information between the first image and the transformation of the second one, without any need for setting control points. A comparison is effectedagainst a publicly available tool, showing the effectiveness of our method.

We present a method to assign to either the foreground or the background the regions into which a gray-level image is partitioned by watershed transformation. Our method is inspired by visual perception in the sense that the border separating any foreground component from the background is detected in correspondence with the locally maximal gray-level changes through the image. The method is implemented as consisting of three steps. The first two steps perform a basic assignment of the regions, while the remaining step examines again some regions tentatively assigned to the background during the second step and possibly changes their status. A feature of the method is that a hierarchical ranking of the regions assigned to the foreground is also accomplished.

The watershed transformation is a primary tool for segmenting a grey-tone image into subsets which are of interest to a visual observer. The resulting image, however, may often appear oversegmented into a large number of tiny regions (basins), most of which are not significant to the problem of domain. In this paper, a method for removing these non-significant basins is presented. The notions of relative significance and intrinsic significance are introduced, which lead to the definition of three types of significance for a basin: strong, weak, and partial. The merging of a basin with other basins only occurs when the significance of the basin is not strong, and is restricted to suitably selected adjacent basins. The merging is performed by using an iterated process consisting of two phases. The first involves with the removal of certain regional minima, and is accomplished by following either a flooding or a digging scheme. The second identifies the basins corresponding to the regional minima remaining in the image and utilizes the watershed transformation. An appropriate selection of the basins to be merged produces a segmented image perceptually close to the original image. The performance of the proposed method is for the case of astronomic images.

Medical semeiotics often deals with patient databases and would greatly benefit from efficient clustering techniques. In this paper a new evolutionary algorithm for data clustering, the Self-sizing Genome Genetic Algorithm, is introduced. It does not use a priori information about the number of clusters. Recombination takes place through a brand-new operator, i.e., gene-pooling, and fitness is based on simultaneously maximizing intra-cluster homogeneity and inter-cluster separability. This algorithm is applied to clustering in dermatological semeiotics. Moreover, a Pathology Addressing Index is defined to quantify utility of the clusters making up a proposed solution in unambiguously addressing towards pathologies.

Most of the classical clustering algorithms are strongly dependent on, and sensitive to, parameters such as number of expected clusters and resolution level. To overcome this drawback, a Genetic Programming framework, capable of performing an automatic data clustering, is presented. Moreover, a novel way of representing clusters which provides intelligible information on patterns is introduced together with an innovative clustering process. The effectiveness of the implemented partitioning system is estimated on a medical domain by means of evaluation indices.

Differential Evolution, a version of an Evolutionary Algorithm, is used to perform automatic classification of handsegmented image parts collected in a seven-class database. Our idea is to exploit it to find the positions of the class centroids in the search space such that for any class the average distance of instances belonging to that class from the relative class centroid is minimized. The performance of the resulting best individual is computed in terms of error rate on the testing set. Then, such a performance is compared against those of other ten classification techniques well known in literature. Results show the effectiveness of the approach in solving the classification task.

Starting from a gray-level image partitioned into regions by watershed segmentation, we introduce a method to assign the regions to the foreground and the background, respectively. The method is inspired by visual perception and identifies the border between foreground and background in correspondence with the locally maximal changes in gray-level. The obtained image representation is hierarchical, both due to the articulation of the assignment process into three steps, aimed at the identification of components of the foreground with decreasing perceptual relevance, and due to a parameter taking into account the distance of each foreground region from the most relevant part in the same foreground component. Foreground components are detected by resorting to both global and local processes. Global assignment, cheaper from a computational point of view, is accomplished as far as this can be safely done. Local assignment takes place in the presence of conflictual decisions.

We introduce a method to reduce oversegmentation in watershed partitioned images, that is based on the use of a multiresolution representation of the input image. The underlying idea is that the most significant components perceived in the highest resolution image will remain identifiable also at lower resolution. Thus, starting from the image at the highest resolution, we first obtain a multiresolution representation by building a resolution pyramid. Then, we identify the seeds for watershed segmentation on the lower resolution pyramid levels and suitably use them to identify the significant seeds in the highest resolution image. This is finally partitioned by watershed segmentation, providing a satisfactory result. Since different lower resolution levels can be used to identify the seeds, we obtain alternative segmentations of the highest resolution image, so that the user can select the preferred level of detail.

Some recently developed tools for grey level image analysis are presented. We describe a new approach to segmentation, based on the watershed transformation and on the analysis of the morphological features of the regions created by the transformation. A skeletonization algorithm is also outlined, able to produce a one-pixel thick skeleton in presence of irreducible sets. Finally, we describe an algorithm to build a pyramid, where the information contents of the image, in terms of shape and topology, is preserved through lower resolution representations.

The watershed transformation is the primary tool of Mathematical Morphology for image segmentation. However, the resulting image often appears oversegmented into a large number of tiny regions (basins), most of which are not significant in the problem of domain. In this paper, a method for removing non significant basins is presented. The notions of relative significance and intrinsic significance are introduced to restrict the merging of a basin to a number of suitably selected adjacent basins. Such a selection allows one to obtain a segmented image perceptually close to the original image. The good performance of the method is shown for the case of astronomic images.

Data mining deals with the problem of discovering novel and interesting knowledge from large amount of data. This problem is often performed heuristically when the extraction of patterns is difficult using standard query mechanisms or classical statistical methods. In this paper a genetic programming framework, capable of performing an automatic discovery of classification rules easily comprehensible by humans, is presented. A comparison with the results achieved by other techniques on a classical benchmark set is carried out. Furthermore, some of the obtained rules are shown and the most discriminating variables are evidenced.

1,5- ICAR-CNR; 2- CNR; 3,4- Dipartimento di Geofisica e Vulcanologia, Universita' di Napoli "Federico II" CNR ISAFOM