ABSTRACT. This paper investigates synthetic data generation through Generative Adversarial Networks (GANs) and Transfer Learning (TL), focusing on Chronic Kidney Diseases (CKD). It analyzes whether GANs, particularly the medGAN and CorGAN architectures, can generate high-quality synthetic tabular data and how TL can enhance this process. The contributions include evaluating alternative medGAN and CorGAN setups, incorporating WGAN and WGAN-GP loss functions, and assessing how fine-tuning with TL impacts data realism and classifier performance. The models were pre-trained on a larger CKD dataset and fine-tuned on a smaller one, using consistent hyperparameters with reduced learning rates during fine-tuning. Experiments involved training Random Forest classifiers in various settings: using real data, synthetic data, and a combination of both, with and without TL. Metrics like dimension-wise probability and multiple training/testing scenarios are employed to assess the quality and utility of the generated data. The synthetic data, especially when generated using TL, improved classifier performance significantly in scenarios where training and testing datasets differed. Notably, F1-scores improved by up to 74.3\% when using TL-generated data. These findings support the use of GANs with TL as a powerful approach to overcome data limitations in healthcare research.

ABSTRACT. Synthetic data generation is an emerging solution to address current data privacy and availability challenges, especially in healthcare applications and wearable devices. This study aims to generate synthetic multivariate time-series data that simulates physical activities, preserving the variability and dependencies observed in real-world data (RWD). An innovative approach that includes data organization, preprocessing, and model training on a dataset of 100 users with several datetime, numeric, and categorical variables is conducted. By using two artificial intelligence (AI) models —Periodic Autoregressive (PAR) and Conditional Tabular Generative Adversarial Network (CTGAN)— a total of 100 new synthetic users with temporal and feature-specific activity patterns were generated. Evaluation results showed strong alignment between synthetic and original data, with mean values of 0.014 for Kolmogorov-Smirnov (KS) Test, 0.057 for Jensen-Shannon (JS) Distance, and 0.011 for Pairwise Correlation, indicating realistic data relationships and feature distributions. Despite these preliminary results, future work includes enhancing computational efficiency and scalability and expanding its generalizability across diverse datasets.

ABSTRACT. Suicide remains a critical public health problem, with increasing cases of Suicide Ideation (SI) necessitating improved identification and intervention strategies. Although generative Large Language Models (LLMs) have demonstrated potential in text analysis for mental health applications, their effectiveness in accurately detecting SI and generating reliable explanations remains underexplored, specially in Brazilian Portuguese (PT-BR) language. This study proposes a new architecture for the Boamente, an AI-based system for suicide prevention. The proposal aims to incorporate into Boamente architecture an explanation generation stage through prompt engineering. The aim is to improve text detection through the classifier model and exploit prompt engineering through an explainer model to generate explanations of why a text in PT-BR does or does not contain SI, increasing the model's interpretability and providing more transparent justifications for its predictions. The proposed structure expands the Boamente architecture by exploring generative LLMs as classifier models (i.e., SI identification) and integrating an explainer model, which generates justifications for predictions. A quantitative evaluation was conducted with different LLMs using classification performance metrics, in which Qwen 2.5 (14B) achieved the highest AUC (0.9898), while the 3B and 7B versions of the same model achieved the best Recall (0.9545). In addition, a qualitative evaluation was conducted with the participation of three professionals (a computer scientist, a linguist, and a psychologist) to analyze the explanations generated by LLMs. The participants considered that LLaMA 3.1 (8B) produced the highest quality justifications. The findings highlight the potential of combining classification and explanation LLMs to enhance explainability and trust in an AI-driven system for suicide prevention.

ABSTRACT. The rapid evolution of artificial intelligence is revolutionizing societies and businesses, with advancements in Natural Language Processing (NLP) models like GPT and BERT enabling novel applications in various fields. This paper introduces a proposal to identify and expand abbreviations in Brazilian Portuguese clinical narratives, addressing a key challenge in interpreting medical texts. Utilizing the SemClinBr dataset, which contains 1000 anonymized clinical texts, the solution consists in a combination of GPT model, prompt engineering and few-shot learning techniques. The result achieved 83.2% success rate in identifying abbreviations and a 94% success rate in correctly expanding them. These results highlight the potential of Large Language Models (LLMs) to facilitate the understanding of medical documents by health specialists from different fields and for applications that use medical narratives.

ABSTRACT. The silent progression of cardiovascular disease (CVD) is a worldwide problem, although new techniques enabled by the rise of electronic health records may aid in reducing its prevalence. Both public health research and big data applications, such as digital twins, are dependent on access to longitudinal and sensitive data; a challenge which may be facilitated by access to longitudinal synthetic data. In this study, we establish a fidelity benchmark for longitudinal synthetic data by extending a well-known method for cross-sectional synthetic data to a longitudinal application within CVD. We find that the univariate distributional difference between the real and the synthetic data is kept low and that pairwise relations are preserved in the synthetic data. Further, we see that the variable-wise temporal trends are preserved, yet may be more extensively studied and have some room for improvement. The study is important to enable future studies within public health prevention and cardiovascular digital twins.

ABSTRACT. While routine macroscopic placental assessment is performed in all deliveries, anatomopathological examination is reserved for selected cases due to logistical and resource constraints. This study explores whether deep learning models analyzing macroscopic photographs could provide preliminary morphological characterization to support clinical decisions regarding anatomopathological examination requests. A dataset of 156 annotated images was expanded to 374 through data augmentation, combining standardized and non-standardized acquisition conditions. Three architectures were evaluated: YOLO11 for instance-level analysis, and ResNet34 U-Net and EfficientNet-B0 U-Net for semantic segmentation. Results showed that YOLO11 performed well on large, well-defined structures, while U-Net–based models effectively characterized broader regions. Standardized image acquisition improved the Dice score by 15%. These findings support the feasibility of incorporating AI-based photographic analysis into routine clinical workflows for placental evaluation.

ABSTRACT. Denoising Diffusion Models (DDMs) are widely used for high-quality image generation and medical image segmentation but often rely on Unet-based architectures, leading to high computational overhead, especially with high-resolution images. This work proposes three NCA-based improvements for diffusion-based medical image segmentation. First, CBAM-MedSegDiffNCA incorporates channel and spatial attention for improved segmentation. Second, Multi-MedSegDiffNCA uses a multilevel NCA framework to refine rough noise estimates generated by lower-level NCA models. Third, MultiCBAM-MedSegDiffNCA combines these methods with a new RGB channel loss for semantic guidance. Evaluations on Lesion segmentation show that MultiCBAM-MedSegDiffNCA matches Unet-based model performance with a dice score of 87.84% while using 60-110 times fewer parameters and 5 times faster training, offering an efficient solution for low-resource medical settings.

ABSTRACT. Video object segmentation is vital for applications like medical diagnostics, but acquiring dense pixel-level annotations, especially for specialized domains like polyp segmentation, remains a major bottleneck. Foundational models offer zero-shot segmentation but typically require manual prompting, which is impractical for long videos. We propose Map2VidSeg, a novel pipeline that automatically generates prompts from image-level classification labels. It leverages localization cues (attention maps/CAMs) from a trained image classifier (ViT/CNN) to create bounding box prompts. These guide an efficient model (YOLOE) with tracking (BOT-SORT) and bidirectional propagation for initial segmentation. Optionally, a high-fidelity model (SAM-2) refines these masks using temporal memory and fusion. Demonstrated on the challenging SUN-SEG benchmark, fine-tuned DINOv2 (ViT) prompts significantly outperform DenseNet-121 (CNN). Our best configuration (DINOv2+YOLOE+SAM-2 Bidirectional) achieves Dice/mIoU 0.76/0.70 (Easy Unseen) and 0.66/0.60 (Hard Unseen), showcasing the viability of robust video segmentation without segmentation training data.

ABSTRACT. Breast cancer remains one of the most prevalent and life-threatening diseases worldwide, being the most frequently diagnosed cancer among women and the second leading cause of cancer-related mortality. Precise tumor segmentation is essential for breast cancer assessment, as it enables accurate estimation of tumor size, monitoring of disease progression, and evaluation of treatment effectiveness. Despite the importance of this task, the development of reliable automatic methods is hindered by the scarcity of fully annotated datasets, which makes manual labeling both time-consuming and subject to inter-observer variability. In this study, we propose an unsupervised 3D tumor segmentation method based on Fuzzy C-Means (FCM) clustering, specifically designed for volumetric Dynamic Contrast-Enhanced MRI (DCE-MRI) of the breast. Unlike supervised deep learning approaches, our method does not require manual annotations for training, making it especially valuable in scenarios with limited labeled data. The proposed pipeline combines prepro- cessing, region-of-interest extraction, and FCM-based clustering to generate accurate segmentation masks with minimal human intervention. We evaluated our approach using clinical data from the ACRIN-6698 dataset, comparing the automatic segmenta- tions against expert manual annotations. The method achieved high performance across multiple metrics, including accuracy, precision, recall, specificity, Dice-Sørensen coefficient (DSC), and Jaccard index (IoU). These results demonstrate the feasibility of unsupervised clustering techniques for volumetric breast tumor segmentation, offering a promising alternative to supervised methods in clinical contexts where annotated data is limited.

ABSTRACT. Breast cancer remains one of the most prevalent and life-threatening diseases among women worldwide. Accurate and early segmentation of tumors in ultrasound images plays a critical role in reducing mortality and guiding treatment. This paper presents a novel deep learning framework that combines an Attention U-Net with a Transformer architecture to enhance segmentation performance. The proposed fused model effectively captures both local spatial features and long-range dependencies, improving delineation of tumor boundaries. In contrast to conventional training paradigms, we propose an incremental learning strategy: the model is first trained on benign tumor images and subsequently fine-tuned on malignant cases. Experimental evaluations on the BUSI dataset demonstrate that our approach significantly outperforms traditional U-Net, TransU-Net, and standalone Attention U-Net models. Notably, the fused model achieves superior performance in segmenting malignant tumors, validating the effectiveness of combining attention mechanisms with global contextual learning through the proposed incremental transfer learning approach.

ABSTRACT. Cardiovascular diseases (CVDs) represent a significant global health concern, necessitating effective early detection methods. The advent of wearable electrocardiogram (ECG) devices offers the potential for enhanced portability and continuous monitoring. However, the substantial volume of data generated underscores the need for automated analysis techniques. This paper presents a framework for intra-subject clustering of ECG heartbeats, employing a robust P-QRS-T wave classifier. A Multiscale Convolutional Neural Network (CNN) in conjunction with a Bidirectional Gated Recurrent Network (bi-GRU) was developed for accurate wave delineation. The methodology involves a pipeline encompassing heartbeat segmentation, feature engineering using both manual descriptors and the Time Series Feature Extraction Library (TSFEL), and subsequent clustering via a k-means algorithm. The proposed wave classifier achieved good overall performance, with 99.68% accuracy in QRS complex identification. Furthermore, a novel ECG database, the Hi Database (HiDB), is introduced. It was acquired using the CompuGroup Medical (CGM) Hi 3-Leads ECG (Hi-ECG), a wearable three-lead device that was commercialised by CGM in collaboration with STMicroelectronics. Intra-subject clustering on the MITDB yielded a mean Adjusted Rand Index (ARI) of 0.78 ± 0.19 and a mean Silhouette Score of 0.65 ± 0.14. Application of the clustering approach to the HiDB resulted in an average Silhouette Score of 0.74 ± 0.13. The findings suggest that the presented framework holds promise for assisting clinicians in ECG annotation tasks and can be adapted to various ECG data sources, particularly those from wearable devices. The intrasubject approach to anomaly detection highlights the potential for personalised cardiac monitoring.

ABSTRACT. Patient stratification—identifying clinically meaningful subgroups—enhances personalized medicine by improving diagnostics and treatment strategies. Electronic Health Records (EHRs) capture rich temporal clinical data recorded throughout hospital care. We leverage four public Intensive Care Unit (ICU) EHR datasets to introduce the first benchmark for evaluating patient stratification over temporal patient repre- sentation learning. We compare statistical methods, LSTMs, and GRUs for generating patient representations and assess their effectiveness in clustering patient trajectories. Using ICD and CCS taxonomies, we propose a hierarchical stratifi- cation benchmark to measure alignment with clinically validated disease groupings. Our results demonstrate that temporal representation learning effectively rediscovers clinically meaningful patient cohorts, achieving state-of-the-art clustering. To further enhance interpretability, we evaluate multiple cluster label assignment strategies. The experiments and benchmark are fully reproducible and available at https://github.com/ds4dh/CBMS 2025 stratification.

ABSTRACT. Chronic Kidney Disease (CDK) is a chronic disease that progressively deteriorates health conditions of the affected patients. It also poses substantial challenges for clinicians in accurately predicting disease progression trajectories and anticipating patient management strategies. This study investigates the potential benefits of using unsupervised clustering techniques as a feature engineering method to enhance the prediction of severe disease trajectories in CKD patients. Specifically, we evaluate the performance of four frequently used clustering algorithms, each with different characteristics and speciality - K-Means, HDBSCAN, Hierarchical Clustering, and Gaussian Mixture Models - on a publicly available clinical dataset (MIMIC IV) formed of detailed records related to CKD patients, including relevant laboratory test results, demographic details, and sequences of ICD-10 codes. The analysis indicates that complex clustering algorithms, like GMM and HDBSCAN, do not have a higher accuracy in detecting clusters on the given clinical dataset than simpler algorithms such as K-Means and Hierarchical clustering.

ABSTRACT. Despite extensive characterization of cellular components within tissues, the fundamental principles governing their organization remain poorly understood. In this study, we propose a comprehensive framework to bridge this gap by analyzing high-resolution 3D images of biological tissues.

Our approach employs a semantic segmentation pipeline to achieve digital segmentation of the primary tissue organelles (cells, nuclei, nucleoli, mitochondrial networks, lipid droplet mass, blood capillaries and hemorrhagic zones). From this, we derive five categories of bioarchitectural parameters: size, shape, inter-organelle distance, orientation, and texture, which together encompass a total of 35 parameters. We then apply unsupervised machine learning techniques to explore the resulting high-dimensional bioarchitectural data, gaining deeper insights into tissue organization.

We validate our framework using patient-derived xenograft (PDX) hepatoblastoma tissue samples acquired through serial block face-scanning electron microscopy. This approach elucidates detailed associations among bioarchitectural parameters and their significance, also enabling the identification of key features that differentiate endothelial cells from tumor cells.

ABSTRACT. Fear of cancer recurrence (FCR) is a significant psychological concern for cancer survivors and is associated with quality of life. While traditional methods for assessing FCR often rely on self-reported questionnaires, these are not always available in clinical settings, and there is no effective measure to reliably identify individuals at risk. This paper proposes two semi-supervised learning (SSL) frameworks to predict FCR using healthcare reimbursement data, which is routinely collected but often underutilized for psychological outcomes such as FCR. Our approach leverages both labeled cases of FCR (patients classified with respect to FCR) and unlabeled cases through a sequential neural network pipeline combined with an autoencoder architecture. The preliminary results demonstrate that SSL can significantly enhance prediction performance, addressing the scarcity of labeled data while offering a scalable framework for real-world healthcare applications, and suggest promising directions for leveraging SSL techniques in predicting high-risk cases based on healthcare reimbursement data. This ongoing work focuses on improving the early identification of high-risk patients, aiming to enable timely interventions and personalized care.

ABSTRACT. Healthcare systems handle thousands of documents daily across various departments, requiring some effort during the digitalization processes. One strategy employed by medical staff is recording appointments for later transcription. However, this process is time-consuming and not practical for all scenarios. In this paper, we present a comprehensive methodology for converting medical audio recordings into structured documentation through multiple AI-based solutions. We propose and evaluate three distinct methods: a baseline two-stage pipeline using Mixtral 7B and Llama 70B models, a cyclic LLM approach leveraging self-improvement loops, and an embedding-based retrieval system utilizing BGE M3. Our experimental results show that the RBF kernel consistently outperformed linear kernels and logistic regression approaches across all metrics, maintaining high precision (0.87-0.94) and perfect recall.

ABSTRACT. Healthcare professionals spend a significant portion of their time on electronic health records documentation, reducing patient interaction time and increasing operational costs. Our solution implements a two-stage pipeline combining voice-to-text transcription using WhisperV3 and text-to-structure conversion through an embedding-based approach. We address critical challenges in medical documentation automation, including specialized vocabulary processing and the prevention of hallucinations in generated content. The system was developed with continuous input from medical domain experts, resulting in a comprehensive field structure covering 78 essential information categories organized into six distinct sections. Our application processes clinical conversations locally, prioritizing data privacy and security while transforming unstructured medical notes into structured clinical documentation. The resulting system enables healthcare professionals to focus more time on patient care while simultaneously improving the quality and accessibility of medical records for better clinical decision-making.

ABSTRACT. The unstructured nature of psychiatric clinical notes poses a significant challenge for automated information extraction and data structuring. In this study, we explore the use of transformer-based language models to perform Named Entity Recognition (NER) on de-identified Spanish electronic health records (EHRs) provided by the Psychiatry Service of Complejo Asistencial Universitario de León (CAULE). A manually annotated gold standard, consisting of 200 clinical notes, was developed by domain experts to evaluate the performance of five models: BETO (cased and uncased), ALBETO, ClinicalBERT, and Bio\_ClinicalBERT. Each model was fine-tuned and assessed using a strict exact matching criterion across six clinically relevant label types. Results demonstrate that ClinicalBERT, despite being pre-trained on English medical corpora, achieved the highest macro-average F1-score on the test set (80\%). However, BETO-cased outperformed ClinicalBERT in four out of six label types, being better in categories with higher syntactic variability. Lower-performing models, such as ALBETO and Bio\_ClinicalBERT, struggled to generalize to Spanish psychiatric language, likely due to domain and language mismatches. This work highlights the effectiveness of transformer-based architectures for structuring psychiatric narratives in Spanish and provides a robust foundation for future clinical NLP applications in non-English contexts.

ABSTRACT. This paper reports preliminary results on the impact of different configurations of clinical text in the quality of the outputs of language models for the task of ICD code assignment. A novel dataset of hospital discharge texts in Spanish extracted over a five-year period from a large healthcare institution is used for the evaluation. The dataset comprises over 142,000 clinical reports, each combining multiple text fields from discharge forms. Different text configurations including full text and specific sections such as, diagnostic orientation, chief complaint and exploration notes. Several pretrained BERT-based models were fine-tuned on these configurations to analyze the effect of context length on classification performance. The results suggest that training BERT models with the full text may not be necessary to achieve good results for ICD Classification.

ABSTRACT. Transfer learning enables the reuse of models trained on large datasets, reducing data collection, computation time, and costs. While widely used in computer vision, its application to models based on electronic health records (EHRs) remains limited. This study evaluates whether fine-tuning an EHR-based model from one country to another outperforms training a model from scratch.

EHR from the SIDIAP (Spain) and CPRD (UK) databases were used, defining a cohort in each country of individuals aged 65+ followed between 2010 and 2019. A prediction model was trained and validated internally for each country to predict 1-year mortality (country-specific), then externally validated and fine-tuned with the other country’s population (re-calibrated model). The models were based on ARIADNEhr, a previously validated architecture. Performance metrics, decision curve analysis, and attention maps were compared.

Participants included 1,456,052 from SIDIAP and 1,507,736 from CPRD, with similar demographics. Performance on the external cohort varied between -10.9% and +39.5%. Fine-tuning consistently improved external performance (1.8%–15.5%), enhanced model calibration and clinical utility, and maintained key contributing variables. However, the fine-tuned models did not reach the performance of the country-specific models, showing a performance drop between 14% and 20%. Fine-tuning may be useful in other fields, but further development may be required for its application in tabular EHR-based prediction models.

ABSTRACT. Automating radiology report generation can significantly reduce the workload of radiologists while improving the accuracy and consistency of clinical documentation. However, achieving optimal alignment between visual and textual representations in medical imaging remains a challenge. To address this, we demonstrate RadGen, a cross-modal fusion based system for automated medical report generation. RadGen uses MedCLIP as both a vision extractor and a retrieval mechanism to enhance the integration of imaging and textual data. By extracting features from retrieved reports and medical images through an attention-based extraction module and integrating them with a fusion module, our system improves the coherence, accuracy, and clinical relevance of generated reports.

ABSTRACT. Breast cancer is one of the most common and lethal types of cancer worldwide, with millions of new cases diagnosed each year. Early detection is pivotal in improving patient outcomes and significantly increases the chances of successful treatment. While traditional detection methods such as mammography are effective, they can be invasive, costly, painful, and less applicable for younger women with denser breast tissue. In this context, infrared thermography emerges as a promising, non-invasive technique for breast cancer detection. However, analyzing these images presents challenges due to noise and irrelevant information that can interfere with accurate diagnosis. Proper segmentation of the region of interest (ROI) is critical to remove these artifacts, focusing on the breast tissue where abnormalities are most likely to occur. In this work, we propose a method for segmenting infrared breast images using the DeepLabV3+ Convolutional Neural Network (CNN), a state-of-the-art architecture known for its effectiveness in semantic image segmentation tasks. Our approach leverages the power of deep learning to precisely delineate breast regions, enabling more accurate feature extraction for subsequent classification tasks. The proposed method was trained and evaluated using a publicly available dataset of infrared breast images, achieving an average accuracy of 98.69\%, an Intersection over Union (IoU) of 97.18\%, and a precision of 98.48\%. These results demonstrate a clear improvement over previous approaches, particularly in terms of segmentation quality, making our method a robust tool for enhancing automated breast cancer detection.

ABSTRACT. Breast cancer remains the second leading cause of cancer-related mortality among women. Early detection and accurate classification of masses and microcalcifications are therefore critical for improving patient outcomes. This study assesses the performance of machine learning (ML) models, specifically Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM), while also evaluating the impact of dataset size variation. The analysis was conducted on three different subsets derived from 1,219 mammograms from the publicly available CBIS-DDSM database using the matRadiomics toolbox. Among the evaluated models, LDA yielded the highest mean area under the curve (AUC) in validation, achieving 68.28% for microcalcifications and 61.53% for masses. These findings underscore the limitations of ML-based radiomics approaches in breast cancer classification and highlight the need to explore more advanced artificial intelligence methods to enhance diagnostic accuracy.

ABSTRACT. Current clinical approaches fail to fully utilize unstructured data in managing prostate cancer (PCa) and kidney cancer (KC), leading to inefficiencies in patient care and increased costs. Effective diagnostics and treatments depend on integrating multimodal data, yet progress is hampered by limited data accessibility and a lack of collaborative validation between clinicians and computer scientists. To address these challenges, the EU-funded COMFORT project aims to develop commercially viable, data-driven multimodal decision support systems. These systems will improve clinical prognostication, patient stratification, and personalized treatment while also assessing the trust that healthcare professionals and patients place in AI-driven tools.

ABSTRACT. Deep Learning approaches show promise for improving tomosynthesis reconstruction but require paired CT and tomosynthesis datasets that are difficult to obtain. This work presents ChestXsim, an open-source Python framework that enable the simulation of digital chest tomosynthesis (DCT) from chest CT data. Its modular design includes the preprocessing pipeline to adapt helical chest CT volumes to a standard tomosynthesis positioning and the simulation of polychromatic projections with noise modelling. Additionally, ChestXsim provides reconstruction techniques (FDK/SART) and leverages GPU acceleration through open-source ASTRA kernels. ChestXsim offers a fast and efficient method for simulating DCT from standard chest CT data, making it a valuable tool for generating paired datasets.

ABSTRACT. Assessing human movement quality is essential in healthcare for diagnosing and monitoring musculoskeletal disorders. Traditional annotation methods are time-consuming and provide limited insights. In this work we present a web-based annotation tool tailored for efficient expert annotation of human exercises, featuring dual-angle RGB video views coupled with interactive 3D visualization of the movement and patient data integration. Through the concurrent streaming of 2D and 3D visualizations of a movement, the platform aims to streamline dataset creation, enhance annotation precision, and support the development of machine-learning based, Functional Capacity Evaluation (FCE) and rehabilitation systems.

ABSTRACT. Concerns about post-surgical care for patients with breast cancer have demanded the development of better biomechanical breast models. However, this machine learning approach is one which requires large amounts of high-quality magnetic resonance imaging (MRI) training data that is of difficult acquisition and availability. This can be solved using synthetic data, with the caveat that generating high resolution images comes at the price of very high computational constraints. As such, generating lower resolution samples is both more efficient and yields better results, even if not to the standards of health professionals. Therefore, this work aims to validate a joint approach between lower resolution generative models and the proposed super-resolution architecture, titled Shifted Window Image Restoration (SWinIR), which was used to achieve a 4x increase in image size of breast cancer patient MRI samples. Results prove to be promising and to further expand upon the super-resolution state-of-the-art, achieving good maximum peak signal-to-noise ratio of 41.36 and structural similarity index values of 0.962 and thus beating traditional methods and other machine learning architectures.

ABSTRACT. Deep learning (DL) electrocardiogram (ECG) predictions are highly valuable for advancing early diagnosis, risk assessment, and treatment planning for cardiovascular diseases. For developing effective predictive models, ECG synthetic data generation is a promising solution for enhancing training datasets while ensuring patient privacy. This work investigates the use of Transformer-based Generative Adversarial Networks (GANs), specifically the TTS-CGAN model, to generate realistic synthetic ECG time-series data. The TTS-CGAN model architecture relies on Transformer encoders for both the generator and discriminator, incorporating state-of-art self-attention mechanisms to enhance time-series data modeling. The research contributions include (i) the analysis of TTS-CGAN for ECG synthesis, (ii) the assessment of synthetic data quality using both MIMIC-IV and MIT-BIH datasets, (iii) the key introduction of hybrid datasets combining real and synthetic ECG signals to assess model training performance, and (iv) the evaluation of generated data using qualitative and quantitative metrics. Experiments demonstrate that synthetic ECG signals maintain statistical properties similar to real data, with high Cosine similarity (up to 0.99) and low divergence (Jensen-Shannon distance < 0.2). Hybrid dataset experiments reveal that models trained on a mix of real and synthetic data (up to 60% synthetic) retain classification performance while improving dataset diversity.

ABSTRACT. Contrast-Enhanced Spectral Mammography (CESM) is a dual-energy mammographic technique that improves lesion visibility through the administration of an iodinated contrast agent. It acquires both a low-energy image, comparable to standard mammography, and a high-energy image, which are then combined to produce a dual-energy subtracted image highlighting lesion contrast enhancement. While CESM offers superior diagnostic accuracy compared to standard mammography, its use entails higher radiation exposure and potential side effects associated with the contrast medium. To address these limitations, we propose GSeg-CycleGAN, a generative deep learning framework for Virtual Contrast Enhancement (VCE) in CESM. The model synthesizes high-fidelity dual-energy subtracted images from low-energy images, leveraging lesion segmentation maps to guide the generative process and improve lesion reconstruction. Building upon the standard CycleGAN architecture, GSeg-CycleGAN introduces localized loss terms focused on lesion areas, enhancing the synthesis of diagnostically relevant regions. Experiments on the CESM@UCBM dataset demonstrate that GSeg-CycleGAN outperforms the baseline in terms of PSNR and SSIM, while maintaining competitive MSE and VIF. Qualitative evaluations further confirm improved lesion fidelity in the generated images. These results suggest that segmentation-aware generative models offer a viable pathway toward contrast-free CESM alternatives.

ABSTRACT. Drug repurposing accelerates drug discovery by identifying new therapeutic uses for existing drugs, but validating computational predictions remains a challenge. Large Language Models (LLMs) offer a potential solution by analyzing biomedical literature to assess drug-disease associations. This study evaluates four LLMs (GPT-4o, Claude-3, Gemini-2, and DeepSeek) using ten prompt strategies to validate repurposing hypotheses. The best-performing prompts and models were tested on 30 pathway-based cases and 10 benchmark cases. Results show that structured prompts enhance LLM accuracy, with GPT-4o and DeepSeek emerging as the most reliable models. Benchmark cases achieved significantly higher accuracy, precision, and F1-score (p < 0.001), while recall remained consistent across datasets. These findings highlight LLMs' potential in drug repurposing validation while emphasizing the need for structured prompts and human oversight.

ABSTRACT. Cancer remains a major global health threat due to its high mortality rate and the challenges associated with its treatment, especially in the later stages. It is very important that the patient identify the cancer as soon as possible to increase recovery chances, and for this, histological images are often used. These images are often looked at by a professional, who analyzes it and categorizes the tissue in labels, but it is often a difficult process for these professionals to analyze a great number of images, and that is why it is used computer vision and neural networks to aid in the identification steps of the disease. A very important network structure that can be used in computer vision is a model called U-Net, named after the U shape made by the decoding and encoding blocks, this network model extracts information while changing the size of the image, being able to get the finest to the more general features of the image. This network can also use an attention system to further improve the feature extraction phase and aid in even better segmentation, with multiple attention models for different purposes. Therefore, this study shows how these attention channels can be tweaked to improve the model results, allowing different types of attention to improve in areas of weakness of the model.

ABSTRACT. Rigorous and periodic monitoring of healing progression in chronic wounds is crucial for the proper management of this impactful health problem; however, manual assessment is highly subjective. Artificial Intelligence-based digital systems have arisen as a solution to automate the analysis of wound properties and reduce the variability of the assessment process at different levels. This study presents an automated approach for wound bed characterisation, using open wound detection and tissue segmentation algorithms to estimate the relative proportion of granulation, slough and eschar tissues. The impact of dataset composition on the performance of deep learning-based tissue segmentation models is investigated, along with the comparison of two model architectures (DeepLabV3+ with a ResNet50 backbone and UPerNet with a Swin Transformer backbone). Two private Wounds datasets are used and augmented with an external third-party external dataset. The results demonstrate that DeepLabV3+ outperforms UPerNet-Swin across all dataset combinations, establishing it as the preferred architecture in this case. Furthermore, incorporating all available datasets improves segmentation performance, underscoring the importance of data diversity. The best-performing model achieved Dice scores of 74.65%, 62.07% and 69% for granulation, slough, and eschar tissues, respectively, with corresponding mean absolute errors (MAE) of 16.51%, 14.68%, and 4.26% for tissue proportion estimation. A comparison of the standard deviation of the obtained MAE results with the ones reported for clinical experts in a similar task demonstrated that the proposed pipeline effectively decreased the variability of the estimated tissue percentages, providing a framework with the potential to streamline the wound monitoring process and increase its reproducibility.

ABSTRACT. Recent advancements in foundation models, such as the Segment Anything Model (SAM), have shown strong performance in various vision tasks, particularly image segmentation, due to their impressive zero-shot segmentation capabilities. However, effectively adapting such models for medical image classification is still a less explored topic. In this paper, we introduce a new framework to adapt SAM for medical image classification. First, we utilize the SAM image encoder as a feature extractor to capture segmentation-based features that convey important spatial and contextual details of the image, while freezing its weights to avoid unnecessary overhead during training. Next, we propose a novel Spatially Localized Channel Attention (SLCA) mechanism to compute spatially localized attention weights for the feature maps. The features extracted from SAM’s image encoder are processed through SLCA to compute attention weights, which are then integrated into deep learning classification models to enhance their focus on spatially relevant or meaningful regions of the image, thus improving classification performance. Experimental results on three public medical image classification datasets demonstrate the effectiveness and data-efficiency of our approach.

ABSTRACT. The aim of this study is to develop an algorithm for automatic segmentation of aortic dissection in contrast-enhanced computed tomography (CTA) images. A2.5D architecture was employed to identify the true and false aortic lumen, major collateral vessels, stent-graft, and false lumen thrombosis in both pre- and post-operative conditions. To address the limited size of the dissection dataset, consisting of 82 CTA scans, a transfer learning approach was applied using a model pre-trained on 336 CTA scans from patients with aneurysmal disease. Preliminary results, evaluated in terms of Dice Score, demon strate promising segmentation performance across the targeted structures. However, collection of further annotated data is necessary to enhance model stability and generalizability. In the future, these segmentations could be leveraged to automate medical image analysis, reducing manual effort and variability while improving the assessment and diagnosis of aortic diseases.

ABSTRACT. Accurate cerebrovascular segmentation is crucial for the early diagnosis and treatment of stroke and aneurysm, both of which pose significant health risks. Time-of-Flight Magnetic Resonance Angiography is widely used for non-invasive vascular assessment, but manual segmentation remains time-consuming, labor-intensive, and highly dependent on the skill level of medical experts. To address this challenge, we propose a fully automated cerebrovascular segmentation framework that integrates conventional voxel-based image analysis with graph-derived vascular features. Our model employs a dual-encoder architecture, where a CNN-based encoder processes MRA images while a second encoder extracts structural vessel information from graph feature-based images. The fusion of these complementary feature representations enhances segmentation accuracy by preserving vessel morphology and improving connectivity. The proposed model was trained and validated on the IXI dataset from COSTA dataset and further evaluated on ADAM and LocH1 datasets. The model achieved a Dice similarity score of 0.8510, Hausdorff distance (HD95) of 3.1150 mm, and average surface distance (ASD) of 0.4646 mm on the IXI dataset, outperforming the conventional 3D UNet model. The proposed model also demonstrated superior performance on external datasets, surpassing the baseline model and proving its generalizability. These results indicate that the proposed model provides a more robust and accurate cerebrovascular segmentation framework, demonstrating its potential for clinical applications.

ABSTRACT. Accurate blood pressure (BP) estimation is critical for cardiovascular health monitoring, with cuffless methods like Pulse Arrival Time (PAT) emerging as non-invasive, continuous alternatives. However, the reliability of PAT-based models for BP estimation remains uncertain due to challenges in model calibration and the complexity of BP regulation. This study investigates the performance of various PAT-based models in estimating systolic (SBP) and diastolic (DBP) BP using the AuroraBP dataset, which contains ambulatory data from 547 participants. The best-performing PAT-based model for SBP was M5-PATS with a mean absolute error (MAE) of 14.23 mmHg. For DBP, it was M2-PATo with a MAE of 9.67 mmHg. However, these results did not meet the accuracy criteria outlined by international standards. Additionally, the baseline model outperformed the M5-PATS model for SBP (MAE: 13.04 mmHg) and achieved comparable performance for DBP (MAE: 9.67). Despite promising results in previous studies, our findings highlight the poor performance of PAT-based models, particularly in dynamic and ambulatory conditions. Future research should explore more advanced approaches, such as deep learning, to capture the complexity of BP variations and provide reliable, continuous BP monitoring in real-life settings.

ABSTRACT. This study presents a comparative statistical analysis of hemodynamic responses in the Action Observation Network (AON) using functional near-infrared spectroscopy (fNIRS) under different motor stimuli and durations. It examines cortical activity in response to 5- and 8-second stimuli, comparing right-hand-clenching and finger-tapping tasks. Ten right-handed healthy participants with no history of brain injury or neurological disorders were recruited. Each completed three experiments, analyzed in two groups: one comparing stimulus durations within the same task (hand-clenching) and the other comparing motor tasks (hand-clenching vs. finger-tapping) using 5-second stimuli. The experiments included movement observation (MO), execution (ME), and motor imagery (MI). Statistical analyses involved t-tests for mean differences, Cohen’s d for effect sizes, and visual inspection of hemodynamic maps. Results showed no significant differences between stimulus durations but revealed moderate differences between motor tasks, with greater hemodynamic responses in gross motor tasks. Additionally, oxygenated hemoglobin (HbO) deactivations were observed during MO, while ME exhibited a higher HbO concentration across all three experiments. These findings provide valuable insights for refining fNIRS-based experimental protocols and optimizing motor task selection in neurological research, supporting its potential in detecting altered hemodynamic patterns in brain-injured individuals and advancing clinical diagnosis and rehabilitation.

ABSTRACT. Technological advances have enabled an increase in data collection. Its use for analysis and creation of models positively impacts decision-making in a wide range of areas. In healthcare, the increase in patient data collection has enabled the creation of models that assist professionals in decision-making, allowing for faster, more accurate and personalized interventions. However, the data collected may contain sensitive information. Even if explicit identifiers are removed, they can be linked to publicly available information, which may violate privacy and result in legal sanctions provided for in regulations governing the use of personal data. As a countermeasure, privacy techniques are used to prove a level of privacy in the publication of data, but they end up impacting the usefulness of the data. Another point to highlight is that the person requesting the data does not participate in the anonymization process, and some manipulations performed on the data may make it unusable for the requester's needs. Managing the trade-off between privacy and confidentiality is challenging since the person publishing the data wants privacy, and the person requesting it wants privacy. Based on this challenge, this article presents a methodology where privacy-preserving data publishing occurs interactively between the publisher and the requester, allowing the requester to specify the manipulations accepted on the data based on the privacy level specified by the publisher. The methodology works with an emphasis on structured tabular data and privacy in relation to record and attribute protection.

ABSTRACT. This study aims to develop and validate an automated artificial intelligence (AI)-driven framework for the detection of pleural plaques (PP) presence from CT scans. To achieve this, an existing pre-trained network for PP segmentation from CT lung scans was integrated into a complete framework designed to detect the presence or absence of PP in individuals, thereby eliminating the need to retrain a large deep learning network. The proposed framework incorporates a novel, lightweight machine learning module that bridges the gap between PP segmentation and presence detection. The proposed framework was evaluated in a cohort of retired workers previously exposed to asbestos. The presence or absence of PP was assessed and compared to binary annotations from CT scans, which were manually labeled by three expert radiologists. The results highlighted the framework’s potential in clinical scenarios where precise localization is unnecessary or where traditional segmentation models struggle due to the presence of small, fine lung structures.

ABSTRACT. Protocol non-adherence in teledermatology, particularly during image acquisition, such as missing rulers or incorrect image orientations, leads to invalid examinations, delayed diagnoses, and increased patient burdens. This study addresses these challenges by evaluating modern neural networks architectures for instance segmentation—Mask2Former (transformer-based) and YOLOv11 (hybrid CNN)— and proposing a novel orientation correction pipeline to improve adherence in two protocols: Approximation (identifying rulers/patient tags) and Panoramic (correcting body orientation errors). Using the Santa Catarina State Telemedicine System dataset (14,238 images for Approximation; 3,692 for Panoramic), models were benchmarked against Mask R-CNN measuring computational efficiency and precision metrics. Results demonstrate YOLOv11’s superiority, achieving state-of-the-art performance with 86.08% AP75, reducing segmentation errors by 14% compared to Mask R-CNN, while maintaining computational efficiency (1.3 GB VRAM, 566–699 ms latency). For panoramic protocol images, our post-processing pipeline mitigated orientation errors by methodically rotating misaligned human masks, improving weighted F-scores from 0.51 to 0.82 and significantly reducing misclassifications between valid and invalid poses. While Mask2Former’s transformer-based design architecture exhibited higher precision, the computational demands (11.7 GB VRAM) hindered deployability in low-resource clinics. The study concludes that hybrid models like YOLOv11 optimally balance segmentation accuracy and operational efficiency, offering actionable insights for real-world clinical implementation. This work bridges AI advancements with clinical pragmatism, offering a framework to enhance protocol adherence, reduce invalid examinations by addressing orientation inconsistencies, and optimize diagnostic workflows in resource-constrained teledermatology.

ABSTRACT. Biomedical text classification requires using data annotated by experts, a costly and time-consuming process. To reduce annotation efforts, Active Learning (AL) arises as an alternative to select the texts considered the most informative from an extensive collection of unlabeled data. On the other hand, biomedical texts are characterized by complex patterns, such as numerical features, abbreviations, and typos, which could be effectively captured using Bag-of-Words (BoW) representation instead of embedding representations if the appropriate features are extracted. In this context, character sequences known as Regular Expressions (RegExes) could be used to generate a feature space representative of the texts. Based on this, we propose analyzing the AL process in traditional classifiers using a Bag-of-RegExes. The performance of BoW classifiers based on Naïve Bayes (NB), Random Forest (RF), eXtreme Gradient Boosting (XGB), and Support Vector Machine (SVM) were evaluated on biomedical texts labeled for obesity, obesity types, and smoking classification problems. The results indicate that classifiers improved their performance using a feature space based on RegExes, especially XGB, which used between 51% to 66% of the total number of training texts. These results highlight the potential of using a Bag-of-RegExes to represent complex patterns, improve interpretability, and outperform state-of-the-art classifiers such as Bidirectional Encoder Representations from Transformers (BERT) and Sentence Transformer Fine-tuning (SetFit) while reducing the need for labeled data.

ABSTRACT. The application of large language models in healthcare presents unique challenges, particularly in non-English contexts where linguistic and cultural nuances significantly impact model effectiveness. Due to the lack of standardized evaluation protocols in less-represented languages, such as Brazilian Portuguese, model performance is often assessed through qualitative analysis, making systematic comparisons impossible. In this work, we introduce a novel benchmark for evaluating medical language models in Brazilian Portuguese, addressing a critical gap in AI assessment for healthcare applications. This benchmark is built upon Brazilian medical aptitude tests spanning 2011-2024, enabling extensive evaluation of both specialist and general large language models. Our findings demonstrate that despite advancements in language model capabilities, significant gaps remain in their ability to reason effectively about medical knowledge in Brazilian Portuguese. This benchmark establishes a proper foundation for evaluating and advancing medical language models in Portuguese, creating a standardized framework to guide development toward more effective, equitable, and culturally appropriate AI systems for healthcare in Brazil.

ABSTRACT. Emergency Medical Dispatch Centres (EMDC) in France are facing a growing number of calls, each requiring a rapid and efficient response, leaving no time for emergency medical dispatchers to code case files. However, this coding is essential for improving patient care through epidemiological analysis and, ultimately, for developing a decision support system. This study explores the use and deployment of natural language processing (NLP) models to automatically code records in French, based on the notes taken by call takers during emergency calls.

Our work investigates fine-tuned biomedical pre-trained encoder models models for classifying highly noisy medical records, demonstrating that these models can successfully accomplish this task with a high F1 score (94\%) without requiring more computationally expensive language models. To ensure the long-term reliability of the models, validation methods specifically addressing the challenges posed by EMDC data were developed. NLP can automate the coding of EMDC call records with a high degree of accuracy. However, the models currently rely solely on the written notes in the files. Future work incorporating call transcripts could further enhance their accuracy.

ABSTRACT. Hepatic encephalopathy (HE) is a serious disease in cirrhotic subjects with unclear pathogenesis and no standardized diagnostic method. As HE affects brain function, it can affect handwriting, demanding coordination, and cognitive task. This study explores the potential correlation between HE, liver cirrhosis, and handwriting. The handwriting data were captured using a graphic pen and tablet, with which the subjects performed eight writing tasks and two drawing tasks. A total of 1965 basic and advanced features from both surface and in-air movements were obtained from these data. Several experiments were performed to analyze the impact of different feature groups on the prediction results. In this study, the weighted k-nearest neighbors feature selection (WkNN-FS) was used. The final XGBoost model was trained using the selected set of 300 features, and achieved an accuracy of 81.96% (95% CI: 79.32 – 91.62) and an AUROC of 87.31% (95% CI: 78.96 – 95.66).

ABSTRACT. Soft tissue sarcomas (STS), such as leiomyosarcoma (LMS) and malignant peripheral nerve sheath tumors (MPNST), are aggressive neoplasms with limited treatment alternatives. Comprehending their biological mechanisms requires the examination of gene expression data, provides insights into transcriptional activity. Gene co-expression networks (GCNs) are essential for elucidating functional gene linkages within datasets, depicting genes as nodes and their interactions as edges. Such methods are typically employed for the discovery of potential biomarkers. However, the computational performance for analysing huge genomic datasets requires High-Performance Computing (HPC) technologies, such as GPGPU and distributed computing models, to improve scalability and efficiency. This study uses HPC-based GCN techniques to uncover potential biomarkers associated with the aggressiveness of LMS and MPNST. Through the comparison of co-expression networks from malignant tumor tissues and their normal or benign counterparts, we identify genes exhibiting differential expression patterns that may facilitate sarcoma growth. As a result, six potential biomarkers were identified in the study. This holistic approach improves our comprehension of STS biology while providing novel tools for biomarker identification and prospective therapeutic targeting.

ABSTRACT. Alzheimer's Disease (AD) is a neurodegenerative disorder characterized by complex, cell-type-specific molecular alterations. This study integrates single-cell RNA sequencing (scRNA-seq) with network-based methodologies to decode transcriptional changes across major brain cell types in AD. Using scRNA-seq data from 432,555 single cells, we constructed Protein-Protein Interaction (PPI) networks specific to each cell type and assessed the differential expression of genes in diseased conditions. Our findings reveal that glutamatergic neurons and inhibitory interneurons exhibit the highest transcriptional dysregulation, while pericytes and endothelial cells show limited changes. The analysis identified significant enrichment of Differentially Expressed Genes (DEGs) within the AD protein module. Network analysis highlights highly connected proteins such as HSPB1, which is implicated in proteostasis, and CXCR4, which is involved in neuroinflammation. Our results underscore the importance of cell-type-specific approaches in AD research, demonstrating that neurons experience more extensive dysregulation, while vascular-associated cells play key roles in maintaining Blood-Brain Barrier (BBB) integrity. These insights emphasize the necessity of tailored therapeutic strategies addressing the heterogeneous molecular landscape of AD.

ABSTRACT. Drug repurposing is an effective strategy to identify novel therapeutic options by leveraging existing drugs with known mechanisms of action and safety profiles. This work introduces a pathway-based drug repurposing approach that combines protein-protein interaction (PPI) networks and transcriptomic data to prioritize candidate drugs. The underlying assumption is that if disease 1 and disease 2 are both associated with the same biological pathway, then drugs known to be effective for disease 2 may be potential treatments for disease 1. To explore these relationships, disease–pathway–drug triplets are constructed. Genes shared between disease 1 and the associated pathway are identified, and their proximity is calculated within the PPI network using a Z-score based on random permutations. Candidate drugs for disease 2 are then analyzed using the Connectivity Map data, assessing whether their up- or down-regulated gene signatures are enriched for the genes contributing to the proximity between disease 1 and the pathway. This integration enables the ranking of candidate drugs based on their potential biological impact on disease 1. Finally, top-ranked drug–disease associations are evaluated through manual curation of the scientific literature to assess existing evidence supporting the proposed repurposing hypothesis.

ABSTRACT. In silico analysis provides valuable insights into studying macromolecules, particularly proteins. Protein structure prediction models, like AlphaFold (AF), offer a cost-effective and time-efficient alternative to traditional methods like X-ray crystallography, NMR spectroscopy, and cryo-EM for determining protein structures. These models are increasingly used in protein-ligand interaction studies, a key aspect of drug discovery. Docking and molecular dynamics simulations facilitate this process, and researchers are continuously developing open-access tools for cavity detection and docking to accelerate protein-ligand interaction studies. However, while many of these tools perform well in specific cases, their strengths and weaknesses in analyzing predicted protein structures remain largely unknown. Therefore, it is crucial to compare docking analyses using experimentally determined protein structures and deep learning-based models. In this study, two well-characterized proteins, dopamine D3 receptor with its ligand ETQ and neprilysin with its ligand sacubitrilat, are used to evaluate docking predictions. The docking tools CB-Dock 2 and COACH-D are applied to both X-ray crystallography-derived structures and five different AF-generated models. The objective is to assess the accuracy of these docking approaches and determine whether this strategy can effectively simulate macromolecular behavior in their microenvironment. By doing so, this study aims to generate new insights and contribute to accelerating research in protein-ligand interactions.

ABSTRACT. Convolutional neural network (CNN) and Transformer-based architectures are two dominant deep learning models for polyp segmentation. However, CNNs have limited capability for modeling long-range dependencies, while Transformers incur quadratic computational complexity. Recently, State Space Models such as Mamba have been recognized as a promising approach for polyp segmentation because they not only model long-range interactions effectively but also maintain linear computational complexity. However, Mamba-based architectures still struggle to capture topological features (e.g., connected components, loops, voids), leading to inaccurate boundary delineation and polyp segmentation. To address these limitations, we propose a new approach called Topo-VM-UNetV2, which encodes topological features into the Mamba-based state-of-the-art polyp segmentation model, VM-UNetV2. Our method consists of two stages: Stage 1: VM-UNetV2 is used to generate probability maps (PMs) for the training and test images, which are then used to compute topology attention maps. Specifically, we first compute persistence diagrams of the PMs, then we generate persistence score maps by assigning persistence values (i.e., the difference between death and birth times) of each topological feature to its birth location, finally we transform persistence scores into attention weights using the sigmoid function. Stage 2: These topology attention maps are integrated into the semantics and detail infusion (SDI) module of VM-UNetV2 to form a topologyguided semantics and detail infusion (Topo-SDI) module for enhancing the segmentation results. Extensive experiments on five public polyp segmentation datasets demonstrate the effectiveness of our proposed method. The code will be made publicly available.

ABSTRACT. Medical image segmentation plays a vital role in numerous applications and has gained significant attention since the introduction of the U-Net model, which enabled convolutional neural networks to achieve high performance with manageable computational costs. Recently, attention mechanisms have emerged as a promising approach to enhance model performance by emphasizing relevant features while suppressing irrelevant ones. This study explores the integration of channel and spatial attention mechanisms into the U-Net architecture, evaluating their impact on segmentation performance and computational cost. Experiments conducted on six public medical imaging datasets demonstrated performance improvements, with Intersection over Union (IoU) gains ranging from 1.62% to 33.66% compared to the original U-Net. These results highlight the potential of attention mechanisms to significantly improve the efficiency and effectiveness of medical image segmentation models.

ABSTRACT. Purpose: Oral health encompasses conditions such as caries, periodontal disease, tooth loss, and oral cancer. Detection of oral conditions in time is crucial for effective prevention and treatment, yet clinical inspections have limitations in identifying hidden or inaccessible lesions. Dental radiography aids in the precise diagnosis but may be prone to misinterpretation. Computational techniques offer potential solutions, including tools for cavity detection and root canal guidance, starting with teeth detection in radiographs. We create a modular teeth instance segmentation model to enhance this step. Methods: We annotated 935 dentomaxilo region bounding boxes on panoramic radiographs and segmented 605 of these images, generating 14,582 tooth polygon annotations. RetinaNet detected dentomaxilo regions and Cascade Mask R-CNN identified individual teeth. Results: Our system achieved satisfactory detection of the dentomaxilo region (92.446 mAP, 0.982 F1 score) and tooth segmentation (79.222 mAP, 0.989 F1 score), exceeding the benchmarks of comparable studies. Conclusions: Our proposal effectively segments teeth on panoramic radiographs, with potential for enhancements such as tooth numbering and caries analysis. Besides aiding diagnosis and supporting dentists, it can expedite epidemiological reports and find utility in forensic medicine.

ABSTRACT. Accurate 3D semantic cell segmentation in volumetric electron microscopy images is crucial for analyzing tissue architecture and plays a key role in studying fundamental biological processes. While current 2D semantic segmentation methods benefiting from numerous pre-trained networks available in the literature are effective, extending these approaches to 3D segmentation remains challenging.

This study introduces a novel automated approach for enhancing 3D semantic cell segmentation by propagating 2D cell delineations and integrating intercellular priors to constrain segmentation within cell boundaries.

We demonstrate the efficacy of our approach for automated 3D semantic cell segmentation using patient-derived xenograft (PDX) hepatoblastoma tissue samples acquired by serial block-face scanning electron microscopy. The proposed method significantly improves 3D semantic cell segmentation, paving the way for bioarchitectural investigations at the cellular level using biological and pathological samples.

ABSTRACT. Standard view cardiac computed tomography (CT) images are of great importance in clinical practice due to the valuable diagnostic information they provide. However, navigating these (3D) medical images to identify such a view can represent a time-consuming process, subject to inter-operator variability. This study presents an approach using a Multiscale Vision Transformer (MViT) to infer the rotation parameters required to obtain standard views and subsequently focus on extracting the left ventricular outflow tract (LVOT) view. Once this view is obtained, the aortic valve is detected and segmented using well-known convolutional neural networks (CNNs), specifically ResNet18 for detection and U-Net for segmentation. The methodology used for its training is detailed, along with the metrics employed to evaluate their performance. Finally, the results obtained for each stage are presented. This approach provides a clear understanding of the performance of each model, from rotation parameter prediction to aortic valve detection and segmentation.

ABSTRACT. Pancreatic cancer remains one of the most aggressive malignancies, with limited survival rates and significant variability in patient outcomes. This study evaluates the performance of three machine learning models (Random Forest, Decision Tree, and XGBoost) in predicting patient survival at 3, 12, and 18 months, using data from the Complejo Asistencial Universitario de León (CAULE) Radiology Department. To systematically analyze the impact of different features on survival prediction, the dataset was structured into seven variable groups (G1-G7), incorporating demographic, clinical, and treatment-related information. To address the inherent class imbalance in survival prediction, an Autoencoder-based synthetic data generation approach was applied, ensuring a balanced distribution of survival and non-survival cases across all timeframes. Hyperparameter tuning was performed, and experimental results indicate that Random Forest and XGBoost achieved comparable performance, both obtaining an accuracy above 81\% at 3 months, 83\% at 12 months, and 88\% at 18 months when trained on Group G7. To enhance model interpretability, SHapley Additive exPlanations (SHAP) was applied to the best-performing model, identifying key factors influencing survival.

ABSTRACT. This paper explores the potential of quantum machine learning for breast cancer detection. We designed a binary classification approach using the BreastMNIST dataset and segmented mass regions derived from the BCDR dataset. A quanvolutional layer is employed as a quantum feature extractor, interfaced with elements of classical neural networks, to enhance the detection of malignant and benign patterns in breast tissue. The hybrid quanvolutional neural network aims to mitigate challenges associated with traditional machine learning models, such as feature sparsity and data imbalance. This architecture employs a simple yet efficient design that integrates the strengths of both quantum computing and classical methods, reducing computational complexity while maintaining performance. Results demonstrate the potential of quanvolutions in diagnostic accuracy, offering a promising framework for integrating quantum computing in medical imaging. This approach provides an optimized solution that balances quantum processing with classical systems for more effective and scalable applications.

ABSTRACT. Breast cancer diagnosis using histopathological images is a challenging task due to the scarcity of annotated medical data, particularly for rare cancer stages. Traditional deep learning models struggle to generalize effectively in such low-data scenarios. To address this problem, we propose a few-shot classification framework for breast histopathological images based on metric-based learning. Our approach leverages Vision Transformers (ViTs) for feature extraction, capturing global contextual information better than conventional Convolutional Neural Networks (CNNs). Additionally, we integrate BLIP-2, a Vision Language Model (VLM), to incorporate manual text prompts and contextual textual descriptions, enhancing the model's interpretability and adaptability. The extracted visual and textual features are fused using a novel feature fusion module, and classified the samples based on cosine distance. We evaluated our approach on BreakHis and BACH datasets, showing its effectiveness in few-shot learning (FSL). Our model achieves 57.12% and around 89% in 5-shot setting, respectively, on the BACH and BreakHis datasets. As the number of support samples increases, performance improves. These findings suggest that combining transformer-based architectures with VLMs enhances the performance of FSL based medical image classification systems.

ABSTRACT. Early diagnosis of breast cancer is crucial for improving survival rates. Radiomics and deep learning (DL) have shown significant potential in assisting radiologists with early cancer detection. This paper aims to critically assess the performance of radiomics, DL, and ensemble techniques in detecting cancer from screening mammograms. Two independent datasets were used: the RSNA 2023 Breast Cancer Detection Challenge (11,913 patients) and a Mexican cohort from the TecSalud dataset (19,400 patients). The ConvNeXt DL model was trained on the RSNA dataset and validated on the TecSalud dataset, while radiomics models were developed using the TecSalud dataset and validated with a leave-one-year-out approach. The ensemble method consistently combined and calibrated predictions using the same methodology. Results showed that the ensemble approach achieved the highest area under the curve (AUC) of 0.87, compared to 0.83 for ConvNeXt and 0.80 for radiomics. In conclusion, ensemble methods combining DL and radiomics predictions significantly enhance breast cancer diagnosis from mammograms.

ABSTRACT. This paper presents a framework for incentivising colorectal cancer (CRC) screening programs from the perspective of policymakers and under the assumption that the citizens participating in the program have misaligned objectives. To do so, it leverages tools from adversarial risk analysis to propose an optimal incentive scheme under uncertainty. The work relies on previous work on modeling CRC risk and optimal screening strategies and provides use cases regarding individual and group-based optimal incentives based on a simple financial scheme.

ABSTRACT. This work tests the implementation of Federated Learning to develop Deep Learning algorithms for glucose prediction in Type 1 Diabetes. A Long-short Term Memory model was implemented to predict glucose levels with a forecasting horizon of 30 minutes. The model was trained using both centralized learning and Federated Learning in parallel, employing different aggregation strategies such as FedAvg, FedMedian, and FedTrimmedAvg. The resulting models were evaluated using metrics recommended for glucose prediction, and the results obtained from both training paradigms were compared. The predictive capabilities of the federated models and the feasibility of adopting the federated paradigm in this specific application were demonstrated.

ABSTRACT. Type 1 diabetes is a chronic disease that results from insufficient insulin production by the pancreas. While artificial pancreas systems have emerged as an alternative therapy, current commercial devices do not account for physical activity in their control algorithms. A major issue is the scarcity of high-quality and unbiased datasets comprising physical activity records. This study investigates the application of transfer learning techniques and physical activity data to enhance glucose prediction. To this end, we propose three distinct methods: a graph topology model, a substitution model and a retraining method. A comprehensive analysis was conducted to assess accuracy, delay and clinical utility. The study revealed that the substitution model exhibited superior accuracy and reduced delay compared to the base model. The study compares the graph topology model and the retrained model, with the former proving to be the best for a prediction horizon of 30 minutes, obtaining a RMSE of 21.63 mg/dL compared to the RMSE of 21.85 mg/dL obtained with the retrained model. For a prediction horizon of 60 minutes, the graph topology model achieved an RMSE of 33.87 mg/dL in comparison to the 34.14 mg/dL of the retrained model. In the Error Grid analisis for a prediction horizon of 30 minutes, both the graph (98.56\%) and the retrained model (98.41\%) achieve results close to the acceptable range. However, with a prediction horizon of 60 minutes, the clinical utility drops significantly (96.4\% vs. 96.21\%). These findings underscore the importance of incorporating physical activity data and the need for further exploration of approaches that account for their impact.

ABSTRACT. Accurate blood glucose prediction is vital for effective diabetes management. However, traditional models focus primarily on improving accuracy, often overlooking the importance of quantifying uncertainty, which is essential for informed clinical decision-making. This study presents an uncertaintyaware blood glucose prediction model using a Bayesian neural network, which not only focus on predictive accuracy but also emphasizes uncertainty analysis. The model is trained on blood glucose data and evaluated using Root Mean Squared Error (RMSE) for accuracy and confidence interval coverage to assess uncertainty calibration. The results show that the model achieves an RMSE of 22.21, indicating strong predictive performance. Moreover, the observed glucose values consistently fall within the predicted ±2σ confidence intervals, demonstrating well-calibrated uncertainty estimates. The low correlation (0.06) between glucose rate of change and uncertainty further suggests that the model performs robustly across both stable and fluctuating glucose levels.

ABSTRACT. There are in-silico models to aid both research and treatment of patients with diabetes. In general, these models consider insulin and carbohydrate content of the ingested food only. By considering complete meal composition, i.e., carbohydrates, protein, fat and fiber, we can substantially improve the metabolic simulation to obtain considerably more precise metabolic responses to meal, drugs and physical activity. To do so, we integrated the SmartCGMS framework for glucose level monitoring and control with the state-of-the-art Sirael virtual metabolic machine, which is capable of simulating metabolic responses to a realistic meal composition. Specifically, we adapted the capability of building SmartCGMS for low-power devices to produce a WebAssembly instead of targeting a low-power device. At diabetes.zcu.cz, there is a live demonstration of our work, which can execute on-line, in a web browser. It demonstrates, how a web browser brings the simulation to patients for education, to improve their outcomes by increasing their motivation via learning. In the wider context, the presented work extends research options for both biomedical engineers, physicians and health care professionals, which educate patients.

ABSTRACT. Sleep deprivation significantly impairs driving performance, increasing accident risk. This preliminary study investigates the effects of fatigue on driving stability and evaluates potential countermeasures using a controlled experimental protocol. A high-fidelity driving simulator and electroencephalographic (EEG) analysis were used to assess neurophysiological and behavioral changes. Two participants completed driving sessions in both rested and sleep-deprived conditions. In the fatigued state, they performed four additional trials under different interventions: airflow, loud music, caffeine, and no external aid. Driving performance was measured using the standard deviation of lane position (SDLP), while EEG recorded changes in brain activity. Results showed increased SDLP in the sleep-deprived state, confirming reduced driving stability. Among interventions, caffeine was the most effective in restoring performance, while airflow and music had moderate effects. EEG data indicated elevated theta (4–7 Hz) and delta (0.5–4 Hz) power, signaling reduced alertness, and decreased beta (13–30 Hz) power, reflecting cognitive decline. This preliminary investigation provides a structured experimental approach for studying fatigue effects and countermeasures. While the small sample limits generalizability, findings support EEG as a valuable tool for drowsiness detection. Future research with larger sample size is needed to refine and validate these results.

ABSTRACT. This paper introduces MNE-RT, a Python package designed for real-time neural feature extraction from magnetoencephalography (MEG) and electroencephalography (EEG) signals in Brain-Computer Interface (BCI) systems. The package incorporates efficient algorithms spanning traditional univariate metrics, such as frequency band power and entropy, to advanced bivariate connectivity measures. It is compatible with various recording systems, enabling the extraction of neural targets from brain signals in real time, with potential applications in enhancing neurofeedback efficacy.

ABSTRACT. Lesion’s diameter is one of the four most important features of a mole. When a lesion exceeds 6 millimeters in diameter, it raises a clinical warning. Most modern dermoscopes are equipped with integrated scales, enabling physicians to accurately estimate the lesion’s size. An accurate size measurement is critical for diagnosing and tracking the growth of skin abnormalities, such as moles or melanoma. In this work, we propose the BQD system automatically capable of measuring a lesion’s dimensions and area using the integrated scale from the dermoscope. We evaluate the system with a test set of 434 images with various types of rulers. The results illustrate the efficacy of the BQD system in accurately measuring and assessing various types of skin lesions, highlighting its potential as a reliable tool in dermatological diagnostics.

ABSTRACT. Microvascular decompression (MVD) is a neurosurgical procedure that relieves nerve compression by repositioning or separating offending blood vessels, effectively reducing pain or spasms. Accurate localization of the compression site is crucial for optimal surgical outcomes, as it enables precise identification and decompression of the offending vessel. While horizontal anatomical relationships are easily identified in the surgical view, compressions occurring along the depth axis are more challenging to discern. In this study, we propose a method to measure accurate intraoperative distances during MVD surgery using Depth-Anything-V2. By leveraging the optical properties of standard imaging equipment in conjunction with the depth estimation model, our method computes precise, absolute distances rather than relying solely on relative measurements, achieving distance estimation errors of less than 2 mm compared to intraoperative and preoperative reference measurements.

ABSTRACT. The increasing demand for accurate and efficient cytopathological diagnostics at early stages highlighted the need for computational solutions of handling real-world clinical variability. This study investigates the feasibility of developing a unified instance segmentation model for cytology, capable of analyzing images from multiple staining protocols without the need for stain-specific pipelines. We evaluate state-of-the-art convolutional and transformer-based models on three distinct cytological stains: Papanicolaou, Feulgen and AgNOR. By comparing models trained on single-stain datasets with those trained on combined multi-stain dataset, we demonstrate that unified training not only maintains high segmentation performance but also improves generalization and operational efficiency. These findings support the development of scalable and consistent AI-assisted diagnostic tools, with the potential to standardize workflows and broaden access to quality pathology services diverse healthcare enviroments.

ABSTRACT. Tumor-educated platelets (TEPs) are circulating blood cells implicated as central players in the systemic and local responses to tumor growth, thus altering their RNA profile. To date, some studies have shown that the TEPs transcriptome can be used for a less invasive cancer diagnosis. The objective of this study is to propose a procedure that can identify a set of key genes with diagnostic value for glioblastoma multiforme (GBM). To identify these key genes, we analyzed TEPs RNA-seq data of healthy subjects and GBM patients from three different public datasets (one of them used as main dataset and the others as test sets). We performed differential expression analysis (DEA) and differential co-expression (DCE) network analysis. Specifically, leveraging the main dataset, we first performed DEA analysis to identify differentially expressed genes (DEGs) and then used these genes to construct and analyze the differential co-expression network. From those networks, we extracted centrality metrics and local clustering coefficient to identify key nodes, hence the more suitable genes for diagnostic purposes. Then we tested those key genes on the other two datasets. Our findings shows that genes identified by betweenness centrality exhibit superior diagnostic power compared to: DEGs, other gene sets identified through other metrics and random sets of differentially expressed genes.

ABSTRACT. Protein sequence alignments are essential for identifying proteins’ shared structural and functional features. Detecting short amino acid sequences, termed patterns, across lung cancer and other related datasets facilitates the identification of relevant features. This study builds on previous findings by exploring proteins that share common patterns already identified. Using sequence matching at 5% and 10% occurrence thresholds, we identified 2,368 and 47 patterns, respectively. To reduce complexity and refine the dataset, shorter patterns from the 10% occurrence streamlined the analysis by isolating highly relevant patterns while reducing redundancy among proteins sharing sequence segments. Subsequent analyses integrated structural predictions for protein folding comparison, enabling the detection of patterns in different proteins and the identification of potential key residues. During cavity detection prediction, some amino acids were inspected in detail to assess their impact on protein function and their relevance in drug-target interactions. These insights were considered during docking studies, focusing on proteins used in treatments with pre-described ligands. By connecting raw sequence data to folding structures and functional features, we identified critical protein cavities that underscore the role of mutations in altering protein behavior and influencing drug-target interactions. These findings highlight protein activity's structural foundations and their importance in understanding cancer biology. By uncovering conserved sequence patterns and their structural implications, this study provides insights into potential biomarkers and therapeutic targets, that could aid in developing more effective cancer treatments.

ABSTRACT. The rapid expansion of genomic and biomedical data has paved the way for constructing complex disease networks, offering new insights into disease mechanisms and therapeutic target identification. Central to these networks are disease modules, which are constructed from seed genes that are prioritized based on their relevance to a given disease. Gene prioritization aims to rank genes based on their strength of association with a disease. Resources like DisGeNET, integrated within DISNET knowledge base, assigns a Gene-Disease Association (GDA) score that reflects the confidence in the association between a gene and a disease. This study investigates how both disease module sizes and GDA scores influence the statistical significance of these modules. By characterizing disease modules, filtering the data based on their GDA score thresholds, and analysing the relationship between their size, GDA score, and significance, potential cutoffs for robust module construction are estimated. Our findings suggest that disease modules filtered to include GDA scores above 0.3 and module sizes greater than 10 tend to be significant. These insights provide guidance for optimal gene prioritization and module selection, ultimately enhancing strategies for target identification and drug repurposing in network medicine.

ABSTRACT. Adaptive proton therapy (APT) is essential for the success of proton therapy by adjusting treatment plans to anatomical changes throughout the treatment. However, on-line adaptation poses significant challenges, particularly due to the limitations of cone-beam CT (CBCT), which lacks the accuracy needed for precise dose calculations. To address this, deep learning-based synthetic CT (sCT) generation from CBCT has emerged as a potential solution. Additionally, accurate segmentation of organs at risk (OARs) is crucial for treatment adaptation, yet existing approaches often treat sCT synthesis and segmentation as separate tasks, leading to potential inconsistencies. This work proposes a multi-task model that simultaneously generates sCT images and segments OARs from CBCT, ensuring anatomical coherence between both outputs. By employing shared feature learning, this approach enhances both image quality and segmentation accuracy. Furthermore, incorporating uncertainty quantification helps assess prediction reliability, addressing a key barrier to clinical adoption. Through this strategy, we aim to enhance the robustness of APT.

ABSTRACT. Cell nuclei segmentation in histopathological images is essential for diagnosing oral epithelial dysplasia, a condition associated with an increased risk of oral cancer. Deep learning models have demonstrated significant potential in this task, but challenges persist due to variations in staining, tissue morphology, and artifacts. This study investigates segmentation models and proposes ensemble approaches to improve instance segmentation in OED histological images. The ensemble integrates diverse segmentation models using different voting rules, with the DC-weighted averaging achieving the best results. The proposed method obtained an accuracy of 94.09% and a Dice coefficient of 0.9461, surpassing individual models and demonstrating significant improvement over individual models. Comparative analysis with literature shows that the ensemble obtained competitive performance across multiple datasets. These results reinforce the potential of ensemble learning to enhance segmentation accuracy, contributing to the development of robust computer-aided diagnosis systems.

ABSTRACT. Accurate lung tumor segmentation is crucial for improving diagnosis, treatment planning, and patient outcomes in oncology. However, the complexity of tumor morphology, size, and location poses significant challenges for automated segmentation. This study presents a comprehensive benchmarking analysis of deep learning-based segmentation models, comparing traditional architectures such as U-Net and DeepLabV3, self-configuring models like nnUNet, and foundation models like MedSAM, and MedSAM 2. Evaluating performance across two lung tumor segmentation datasets, we assess segmentation accuracy and computational efficiency under various learning paradigms, including few-shot learning and fine-tuning. The results reveal that while traditional models struggle with tumor delineation, foundation models, particularly MedSAM 2, outperform them in both accuracy and computational efficiency. These findings underscore the potential of foundation models for lung tumor segmentation, highlighting their applicability in improving clinical workflows and patient outcomes.

ABSTRACT. This paper proposes the definition of core components of the healthcare domain that can be used as basic building blocks for the development of mHealth applications. Using Natural Language Processing techniques, we systematically analyze all health interventions defined in the Nursing Interventions Classification taxonomy to define a comprehensive set of basic components. Our results show that it is feasible to define a finite set of core components from the nursing domain language that can be further composed into health care plans thus establishing a foundation for developing mHealth solutions with reduced technical effort.

ABSTRACT. Mobile medical applications provide a variety of functionalities for users, from managing critical personal data to providing basic medical information. However, due to the variety of functionalities and lack of consistent and concrete regulatory oversight across app marketplaces, medical apps potentially pose a threat to users who are generally unaware of app capabilities. Therefore, in order to help legal experts quickly identify which regulatory body applies to medical apps used by consumers, we present a method to convert and plot the prose of both app descriptions and regulatory legalese into a vector space to facilitate rapid cosine similarity scoring. Our study demonstrates how to automate the regulation of mobile medical apps using descriptions in the language of regulatory bodies. Our results show a need for comprehensive regulatory oversight of medical apps, with 54.8% of apps on Google Play and 58% of apps on the Apple App Store.

ABSTRACT. We consider the use of machine learning for convenient, inexpensive monitoring of heart health status using a wrist-worn photoplethysmography (PPG) sensor. Patients' health status is assessed via a standard questionnaire-based measure of symptom severity previously shown to provide information relevant to survival. In a pilot study yielding a nearly class-balanced collection of 198 data records from 15 participants who have been diagnosed with heart failure (heart insufficiency), we attain a macro-averaged classification accuracy of 0.73 using a random forest classifier to predict binarized symptom severity from heart rate variability (HRV) and physical activity metrics extracted from the PPG signals. Our results support the possibility of reducing hospital readmission rates through in-home monitoring, while limiting the frequency of inconvenient and expensive in-person assessments in a clinical setting.

ABSTRACT. Depression is a serious mental health problem worldwide and requires quick and accurate diagnosis. Recently, machine learning and deep learning techniques have been actively applied to depression diagnosis research, especially as audio signals are attracting attention as non-invasive and economical methods. This study proposes the Long-Term Dilated-Conformer (LTD-Conformer), an extension of the existing Conformer model designed to utilize audio signals for more accurate depression detection. The LTD-Conformer employs dilated depthwise convolution to achieve a wide receptive field and integrates a Long-Term Module to capture sequential information in audio features. This model comprehensively captures and analyzes the local, global, and sequential patterns in audio signals. In addition, we combined listening (Mel-spectrogram) features and speaking (HuBERT) features to effectively analyze both perspectives of audio signal. The experiment was conducted using DAIC-WOZ dataset, and the LTD-Conformer model achieved an accuracy of 87.04% and an F1-score of 0.87, demonstrating a 4% improvement in accuracy and a 0.04 increase in the F1 score compared to the existing Conformer model. This study presents the possibility that the audio signal-based depression LTD-Conformer model can be effectively applied to depression diagnosis and will develop into a strong audio-based depression diagnosis model in the future.

ABSTRACT. Hereditary ataxias can lead to the loss of speaking abilities, which in turn can strongly affect a person’s social isolation, and ability to complete daily tasks. A recently developed therapeutic intervention for dysarthria – the loss of speech due to neurological factors - is a therapy program of vocal exercises provided by expert speech pathologists. Due to the repeat and intensive frequency of these exercises, combined with the requirement of close clinical oversight, a novel method of delivery is required. The delivery must allow patients to perform these exercises in a systematic way, provide feedback to their supporting clinician in a timely manner, and be provided with graphical feedback of their progress in real-time both during the exercises and over the longer term. The SpeechAtax project has developed a tablet-based application that supports all of these requirements through novel audio-visual aids, notifications, and text information tailored to the patient’s condition and the stage of progress in their therapy. This paper presents the system architecture of the application, conceptual challenges and solutions, and shows how the platform supports qualitative feedback from a representative population.

ABSTRACT. This work investigates the impact of automatic data augmentation on the generalization performance of CNN models on speech-based Parkinson’s disease classification. We propose a method that sequentially applies 12 voice-specific augmentations, selecting the most effective one based on performance. We use a pre-trained CNN as a feature extractor. To assess inter-dataset generalization, we conduct experiments where each dataset is used for training while the others are used for external validation. The results demonstrate that augmentation helps reduce the generalization gap, with specific augmentation strategies enhancing the accuracy by as much as 25% compared to the baseline setup. Despite the increase in computational complexity due to the proposed method, this study reinforces the importance of augmentation in domain adaptation for speech-based PD classification.