Accurate quantification of cervical spinal cord (CSC) atrophy is crucial for monitoring disease progression in multiple sclerosis (MS). However, clinical magnetic resonance imaging (MRI) protocols are typically limited by low resolution (LR) and inter-slice gaps, impairing reliable volumetric assessment. To address these constraints, we introduced a novel three-dimensional (3D) deep learning (DL) super-resolution (SR) model designed to reconstruct high-resolution (HR) CSC volumes from LR, sparsely sampled MRI scans. The model was evaluated on both public and clinical datasets using standard quantitative metrics, and its impact on segmentation performance was further assessed. We applied the model to both longitudinal and cross-sectional MS datasets comprising, respectively, five subjects with multiple follow-up scans over 4 to 6 years, and 23 subjects with a single scan. The 3D spinal cord super-resolution (SCSR) model consistently generated reliable reconstructions with enhanced anatomical details. This approach improves the reliability of the segmentation in initially challenging clinical LR scans by generating high-quality CSC SR images, allowing for more precise quantification of CSC volume. Finally, longitudinal regional and global CSC atrophy were estimated and compared with healthy controls (HCs), providing preliminary insights that will benefit from further validation in larger cohorts.

Precise measurement of Spinal Cord Cross-Sectional Area (SC-CSA) is crucial for monitoring SC health and diagnosing neurodegenerative diseases. Accurate measurement of SC-CSA is challenging due to several factors such as SC curvature, patient positioning, segmentation methods, and limitations of Magnetic Resonance Imaging (MRI) resolution. Traditional methods that rely on vertebral levels as a reference suffer from inter-subject variability and limited reproducibility. In this study, we propose a novel landmark-based method to define robust anatomical references for standardizing the delineation of SC segment boundaries for SC-CSA measurements. This approach relies on consecutive spinal nerve midpoints (CNMPs) as anatomical references, providing more precise segment delineation and consistent SC-CSA measurements. We validated the effectiveness and robustness of our method using two public datasets of healthy individuals: the first comprises MRI images from different scanners and centers, and the second offers three different head positions for the cervical SC. Experimental results show an intra-subject mean coefficient of variation across head positions of 4.01 ± 2.89

Skin cancer detection is a crucial task in dermatology, with accurate classification of skin lesions playing a key role in improving patient outcomes. This paper proposes a modified architecture based on Modified Multi-Head Attention Transformer (MMHAT) for skin image classification, specifically targeting the detection of malignant melanoma and melanocytic nevus. In this work, we introduce a hybrid approach that integrates convolutions with linear projections, combining the ability of convolutions to effectively capture local spatial features with the capacity of linear projections to model global relationships. This adaptive mechanism enhances the model’s ability to extract both fine-grained and large-scale patterns, leading to a more expressive and robust feature representation. Additionally, we incorporate a ReLU activation function in the adaptive weighting mechanism to ensure that only positive contributions from different components are retained. The hybrid approach is evaluated on the publicly available ISIC 2018 dataset, containing a diverse range of dermoscopic images. Experimental results show that the proposed Transformer model achieves an accuracy of 0.91, outperforming conventional models such as ResNet-50 and InceptionV3, which achieved accuracy of 0.90 and 0.85, respectively. Furthermore, experiments with and without multi-head attention demonstrate the significant impact of the attention mechanism in improving model performance, particularly in skin cancer detection.

The cross-sectional area (CSA) of the cervical spinal cord (CSC) is a key biomarker for monitoring neurodegenerative diseases such as multiple sclerosis (MS). However, its clinical utility is limited by inter-subject variability (I-SV) due to anthropometric factors (height, sex) and MRI scanner differences, which can obscure true anatomical changes. This motivates the development of a normalization strategy to enhance the sensitivity of this established biomarker to pathological changes. Existing normalization methods rely on complex imaging features unavailable in many clinical datasets, thereby motivating the need for simpler approaches. We propose a lightweight regression-based normalization model for CSC-CSA using routinely available predictors. We evaluated the effect of these predictors across cervical spinal levels (CSL), demonstrating that the model captures a quantifiable proportion of I-SV reflected by a 15–23 R^2 of 0.09–0.17. This approach offers a practical, generalizable framework for normalizing CSC-CSA, enabling accurate cross-sectional and longitudinal analyses in MS cohorts, as well as quantifying atrophy linked to progressive MS.

In recent years, Implicit Neural Representations (INRs) have emerged as a powerful and widely studied technique, offering promising results in areas such as 3D scene reconstruction from 2D images (NeRF). They have also shown significant potential in solving various inverse problems, including deblurring, super-resolution, and image reconstruction. This paper investigates the capacity of INRs for non-blind image deconvolution, operating without the need for any external training datasets. State-of-the art blind or semi-blind methods achieve suboptimal results on non-sparse smoothing kernels widely encountered in medical imaging. We propose a self-supervised framework for this task, where the kernel and the image are estimated separatelly. For the blur kernel we employ a compact network and Fourier features encoding, whereas for the image we use hash encoding, better suited for reproducing sharp details. To facilitate comparisons, we are conducting this study on natural images convolved with Gaussian filters. The implemented framework demonstrates promising results for both image and kernel deconvolution, outperforming several existing self-supervised methods.

Skin cancer represents a major global health concern, driving significant research efforts toward improving diagnostic accuracy and treatment outcomes. Early detection remains essential for effective management, as it considerably enhances the chances of successful intervention. With the rising incidence of skin cancer worldwide, there is a critical need for reliable and interpretable diagnostic systems that can support clinicians in making timely and accurate decisions. This study investigates the application of deep learning (DL) models for skin lesion classification, focusing on three convolutional neural network (CNN) architectures: VGG16, DenseNet169, and DenseNet201. The models are trained to classify lesions as benign or malignant, aiming to strengthen early detection capabilities. Comparative analysis shows that all models deliver competitive results, with DenseNet201 achieving the best performance, recording an accuracy of 0.8788, a precision of 0.8776, a recall of 0.8781, and an F1 score of 0.8778. These results surpass several state-of-the-art methods, highlighting DenseNet201 as a robust approach for skin lesion classification. To improve the reliability and interpretability of the outcomes, the Integrated Gradients (IG) method is employed to explain the models’ decision-making process. This explainable artificial intelligence (XAI) technique identifies the most influential features contributing to classification, thereby increasing trust in the predictions and supporting evidence-based clinical practice. The findings emphasize the potential of CNN models, particularly DenseNet201, to advance computer-aided diagnosis (CAD) of skin cancer. By combining high predictive performance with interpretability, such models offer valuable tools to assist healthcare professionals in the early and accurate detection of malignant lesions.

Diabetic retinopathy (DR) is a leading cause of vision loss in individuals with diabetes, making early and accurate detection essential for preventing severe complications. Automated classification of DR stages from retinal fundus images can assist clinicians in timely diagnosis and management. This study proposes a hybrid approach that combines deep learning-based feature extraction with traditional machine learning classifiers for automatic DR stage classification. Two datasets were used: Dataset A (DiabeticRetinopathy_Messidor_EyePACS_Preprocessed) and Dataset B (APTOS 2019 Blindness Detection). Three hybrid architectures were evaluated: MobileNetV2 with Support Vector Machine (SVM), MobileNetV2 with Random Forest (RF), and VGG16 with SVM. The models exploit convolutional neural networks for extracting discriminative features and employ conventional classifiers for robust decision-making. Grad-CAM and Score-CAM techniques were applied to enhance interpretability by visualizing the regions influencing model predictions. Experimental results on Dataset B show that MobileNetV2 + SVM achieved an accuracy of 85

Skin cancer detection is a crucial task in dermatology, with accurate classification of skin lesions playing a key role in improving patient outcomes. This paper proposes a modified architecture based on Multi-Head Attention Transformer (MMHAT) for skin image classification, specifically targeting the detection of malignant melanoma and melanocytic nevus. The Transformer model leverages self-attention mechanisms to capture complex spatial relationships and contextual information across different regions of skin images, which is essential for distinguishing between various types of lesions. Unlike conventional Convolutional Neural Networks (CNNs), which are limited by kernel size and layer depth, the Transformer’s self-attention mechanism enables it to capture long-range dependencies and reduce reliance on external information, allowing for better internal data correlation. In this work, we utilize a multi-head self-attention module to extract image features, which allows the model to focus on multiple aspects of the image simultaneously, enhancing its ability to detect subtle patterns often missed by traditional CNNs. The model is evaluated on the publicly available ISIC 2018 dataset, containing a diverse range of dermoscopic images. Experimental results show that the proposed Transformer model achieves an accuracy of 0.91, outperforming conventional models such as ResNet50 and Inception V3, which achieved accuracy of 0.90 and 0.84, respectively. Furthermore, experiments with and without multi-head attention demonstrate the significant impact of the attention mechanism in improving model performance, particularly in skin cancer detection.

Skin cancer continues to be a significant worldwide health issue, highlighting the critical need for precise and timely detection to guarantee positive patient outcomes and efficient treatment. While deep learning algorithms have demonstrated significant potential, they are frequently observed as black-box models, posing challenges for dermatologists in interpreting and validating their decisions. This study integrates deep learning with explainable artificial intelligence to address the complexities inherent in skin cancer detection. Four advanced pre-trained models—InceptionV3, Xception, ResNet50V2, and DenseNet121—are employed for the classification of skin lesions. To deal with the class imbalance and to improve the generalisation of the model, image augmentation techniques are considered. Transparency in the decision-making process is achieved through XAI, which is essential in medical contexts where interpretability fosters trust and supports the seamless adoption of AI-driven diagnostic systems in clinical workflows. Extensive evaluation reveals that better performances are obtained when Xception model is considered, achieving an accuracy of 90.15

Accurate skin cancer classification is essential for early diagnosis and effective treatment planning, enabling timely interventions and improved patient outcomes. In this paper, the performance of four pre-trained models—two convolutional neural networks (ResNet50 and VGG19) and two vision transformers (ViT-b16 and ViT-b32)—is evaluated in distinguishing malignant from benign skin cancers using a publicly available dermoscopic dataset. Among these models, ResNet50 achieved the highest performance across all the evaluation metrics, with accuracy, precision, and recall of 89.09% and an F1 score of 89.08%, demonstrating its ability to effectively capture complex patterns in skin lesion images. While the other models produced competitive results, ResNet50 exhibited superior robustness and consistency. To enhance model interpretability, two eXplainable Artificial Intelligence (XAI) techniques, Local Interpretable Model-Agnostic Explanations (LIME) and integrated gradients, were employed to provide insights into the decision-making process, fostering trust in automated diagnostic systems. These findings underscore the potential of deep learning for automated skin cancer classification and highlight the importance of model transparency for clinical adoption. As AI technology continues to evolve, its integration into clinical workflows could improve diagnostic accuracy, reduce the workload of healthcare professionals, and enhance patient outcomes.

The Dirichlet Process Mixture Model (DPMM) is a Bayesian nonparametric approach commonly used in unsupervised learning, notable for its ability to infer the number of components in the data. In this work, DPMM is integrated with a Markov Random Field (MRF) to tackle the segmentation of Polarimetric Synthetic Aperture Radar (PolSAR) images. The MRF incorporates spatial context, enhancing segmentation accuracy. Class labels are updated using the Expectation Maximization algorithm. The proposed EM-DPMM-MRF model is evaluated on both simulated and real PolSAR images with known ground truth. The experimental results demonstrate strong and consistent performance.

Epilepsy is a chronic neurological disorder characterized by recurrent clinical episodes stemming from various causes, which lead to significant physical, psychological, social, and intellectual challenges. In recent years, advancements in medical diagnostic techniques for epilepsy have become crucial for mitigating the risks associated with the disorder. Electroencephalography (EEG) remains the gold-standard method for seizure analysis as it captures changes in brain activity. However, individual variability in EEG signals, influenced by factors like age and alertness, makes the visual identification of epileptic abnormalities a complex, time-consuming, and error-prone task. This emphasizes the necessity for automated approaches to accurately analyze EEG signals. In this study, deep learning models—namely Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM)—were utilized to enhance seizure detection accuracy. Data from the UCI Center for Machine Learning and Intelligent Systems, comprising five classes, were employed for two experiments. Experiment 1 involved binary classification of classes 1 (epileptic) and 3 (non-epileptic). Experiment 2 aimed to distinguish between epileptic (class 1) and non-epileptic seizures by combining classes 2, 3, 4, and 5 into a non-epileptic category. The BiLSTM model demonstrated superior performance, achieving 98.00

Skin cancer remains one of the most prevalent malignancies worldwide, with melanoma accounting for the most lethal form due to its high metastatic potential. Early and accurate diagnosis is essential to improving patient survival, yet access to specialized dermatological expertise is limited in many regions. Recent advances in deep learning, particularly Convolutional Neural Networks (CNNs), have significantly enhanced the capabilities of computer-aided diagnosis (CAD) systems. This study introduces and evaluates three lightweight and optimized CNN-based architectures for binary skin cancer classification: (1) a Modified MobileNet with Residual Blocks, (2) an AlexNet enhanced with Squeeze-and-Excitation (SE) Attention, and (3) a custom-designed CNN with integrated Residual Connections. Using a benchmark dermoscopic dataset from the ISIC Archive, we apply standardized preprocessing and data augmentation techniques, followed by rigorous model training and evaluation. Results show that the Modified CNN achieves the highest accuracy (84.70%), precision (84.56%), recall (84.78%), and F1-score (84.63%), outperforming or matching state-of-the-art models such as ResNet-101, while maintaining computational efficiency. These findings support the feasibility of deploying such models in mobile health applications, offering a scalable solution for early melanoma screening in resource-constrained environments.

Skin cancer is a common cancer, and early detection is vital for effective treatment. Automated skin cancer segmentation, especially through deep learning models, has gained significant attention to improve diagnostic accuracy. This paper explored the application of deep learning techniques, specifically transformer and CNN models, for skin lesion segmentation. Two models, U-Net and hybrid ViT, were used with the publicly available PH2 dataset. The models were evaluated based on their ability to segment skin lesions from complex dermoscopic images, particularly when dealing with small datasets. Standard segmentation metrics, including dice coefficient, Intersection over Union (IoU), and accuracy, were used for assessment. To enhance transparency and trust, the Grad-CAM method was applied for explainable AI (XAI). The implemented models, U-Net and hybrid ViT, produced promising results. The hybrid ViT model achieved an accuracy of $96.31 \%$, a dice of $96.01 \%$, and an IoU of $92.33 \%$, whereas the U-Net attained an accuracy of $95.45 \%$, a dice of $94.90 \%$, and an IoU of $90.30 \%$. The study emphasized the effectiveness of the utilized models in accurately segmenting skin lesions, which can enhance diagnostic precision and treatment planning.

Non-invasive fetal electrocardiography (fECG) monitoring is crucial for prenatal care, enabling early detection of fetal cardiac anomalies. However, extracting the weak fECG signal from composite abdominal ECG (aECG) recordings is challenging due to maternal ECG dominance and noise interferences. This paper proposes W-NET, a novel deep learning architecture inspired by one-dimensional U-Net and Transformer models, designed for simultaneous extraction of maternal (mECG) and fetal (fECG) ECG signals from a singlechannel aECG. The model employs a dual-branch structure with adaptive maternal feature suppression in the fetal branch to enhance separation fidelity. Trained on synthetic and real datasets, W_Net outperforms baselines in QRS detection, achieving F1-score 1.000 on synthetic data and 0.824 on real data. To enable practical deployment, we integrate W-NET with a secure Flask-based web interface for real-time signal analysis, visualization, and fetal heart rate classification, thereby facilitating accessible and continuous prenatal monitoring.

Epilepsy is one of the most prevalent neurological disorders and is characterized by abnormal electrical discharges in groups of neurons, leading to recurrent seizures. As a major public health concern, epilepsy can be life-threatening due to the sudden loss of motor control and/or consciousness during seizure episodes. Consequently, accurate and timely seizure detection is essential to reduce patient anxiety, enhance safety, and support effective clinical decision-making. Electroencephalography (EEG) is widely used for monitoring brain activity and remains one of the most reliable tools for seizure detection and treatment optimization. However, manual analysis of EEG recordings is time-consuming, subjective, and prone to human error, highlighting the need for automated and reliable detection systems. In this study, a comparative analysis of six machine learning algorithms is conducted for automatic epileptic seizure detection using EEG data obtained from the UCI Machine Learning Repository. The evaluated classifiers include K-Nearest Neighbors (KNN), Decision Tree, Random Forest (RF), Multilayer Perceptron (MLP) neural network, Extreme Gradient Boosting (XGBoost), and Naive Bayes. Standard preprocessing and feature representation techniques are applied prior to classification. The performance of each model is assessed using commonly adopted evaluation metrics, including accuracy, specificity, precision, recall, and F1 score. Experimental results indicate that the Random Forest classifier achieves superior performance compared to the other models, attaining an accuracy of 97.96%, specificity of 98.86%, precision of 95.43%, recall of 94.41%, and an F1 score of 94.92%. These findings demonstrate the effectiveness and robustness of ensemble-based machine learning methods for EEG-based epileptic seizure detection, suggesting their potential for real-world clinical applications.

Skin cancers represent a significant health issue, posing considerable challenges in early detection and accurate classification. This study investigated the effectiveness of Deep Learning (DL) models for skin cancers classification using dermoscopic images. Two pre-trained models, DenseNet169 and VGG16, were employed. The DenseNet169 model delivered the highest accuracy, reaching 90.61% while, the VGG16 model achieved an accuracy of 84.24%. The two models were assessed using key performance metrics to conduct a comprehensive evaluation of their effectiveness and overall performance. This analysis provided significant insights on various aspects of each model, allowing a detailed comparison. It also provided a clearer comprehension of the strengths and weaknesses of both models, aiding in informed decision-making and driving further optimization. In addition, three explainable AI (XAI) techniques were applied to deliver perceptions into the decision-making process. These techniques enhance model interpretability by emphasizing the regions of the input images that contribute most to the classification decision, thus improving the transparency and reliability of the model's predictions. Reliable skin cancer classification has the potential to assist healthcare specialists, particularly dermatologists, in making early diagnoses, facilitating timely interventions, and ultimately improving patient outcomes. This study offers valuable contributions for enhancing skin cancer detection and classification, supporting accurate diagnosis and treatment development for medical practitioners.

Precise measurement of Spinal Cord Cross-Sectional Area (SC-CSA) is crucial for monitoring SC health state and diagnosing neurodegenerative diseases such as Multiple Sclerosis (MS), Devic Neuromyelitis Optica (DNMO), and Amyotrophic Lateral Sclerosis (ALS) progression. However, accurate measurement of SC-CSA is challenging due to many factors such as SC curvature, patient positioning, segmentation methods, and the limitations of Magnetic Resonance Imaging (MRI) resolution. Traditional methods that rely on vertebral levels as reference suffer from intersubject variability and limited reproducibility. In this study, we propose a novel landmarks-based method for automated SC-CSA measurement. The proposed approach, relying on consecutive spinal nerves’ midpoints as anatomical reference, offers more precise spinal segment delimitation and consistent SC-CSA measurements. We validated the effectiveness and robustness of our method using two public datasets comprising MR Images of healthy individuals covering all cervical levels and for three different head positions. Experimental results demonstrated that the proposed SC-CSA measurement method could be considered as a useful tool for potential clinical applications

Multiple Sclerosis (MS) is a chronic neurological disease of the central nervous system, marked by inflammation, demyelination and scarring (sclerosis) in the brain and spinal cord, resulting in a variety of motor, sensory and cognitive symptoms. This article examines recent advances and future prospects in the in-depth exploration of Deep Learning (DL) techniques for segmenting MS lesions. Highlighting recent developments, it explores innovative methods for lesion identification and segmentation, while addressing persistent challenges. This review offers a critical analysis of the current research landscape and suggests promising avenues for the future evolution of these techniques.

Purpose: Multiple Sclerosis (MS) is a chronic disease of the Central Nervous System (CNS), characterized by the presence of disseminated lesions in the brain and Spinal Cord (SC). Magnetic Resonance Imaging (MRI) has become an essential tool for studying the anatomy and functions of the CNS in vivo, enabling not only the identification of brain structures but also the detection of damaged tissue in various neurodegenerative diseases, including MS. The segmentation of lesions on MR images is a crucial step in the diagnosis and monitoring of the disease. However, manual segmentation of MS lesions is a complex and time-consuming task requiring considerable expertise. Methods: This paper proposes a fully automated method for MS lesion segmentation based on a Convolutional Neural Network (CNN) architecture. The model was trained on datasets from the MICCAI 2016 and ISBI 2015 international challenges. FLAIR images from these databases were used as input to the CNN. Results: The results show a significant improvement in the accuracy and robustness of the model, resulting in high-quality segmentation of MS lesions. The model achieved remarkable performance, with a Dice Similarity Coefficient (DSC) of over 89%, outperforming recent methods. Conclusion: These promising results underline the considerable potential for future advances in the automated segmentation of MS lesions.