ABSTRACT. Irregular sleep activity is a well-known symptom of dementia, while assessing changes in circadian rhythm activity in clinical settings remains challenging. In this study, we develop a multi-harmonic visualization model for circadian rhythm from actigraph measurements. By applying an evolutionary learning fitting procedure, we obtain models that provide more detailed information about circadian rhythm activity. The results show that the proposed model is able to reflect the alterations in circadian rhythm that correspond to the severity of cognitive impairment levels and sleep disturbances. The visualization thus shows nuanced changes in circadian activity over time, surpassing the capabilities of standard models. This approach can provide valuable insights for the care of people with dementia, both during the progression of symptoms and at the end of life.

ABSTRACT. This study proposes a classification framework for Alzheimer’s Disease (AD) using neuroimaging to detect structural brain alterations. It integrates multi-region analysis with displacement vector fields and Jacobian determinants. Magnetic resonance images from 432 cognitively normal (CN), 341 mild cognitive impairment (MCI), and 245 AD individuals were analyzed. Groupwise registration and deformable coregistration quantified spatial deformations and local volumetric changes. Statistical moments from displacement vector fields and Jacobian determinants enabled region-specific analysis. Stratified by sex, CN vs. AD classification achieved an AUC of 0.93 and 88.68% accuracy for males, and an AUC of 0.94 with 87.89% accuracy for females, demonstrating the efficacy of deformation-based biomarkers for AD diagnosis.

ABSTRACT. Anthropometric measurements are used for numerous biomedical applications, such as the management of dietary, medication, and orthopedic treatments. However, for some patients, these measures can be impaired due to the high complexity of their clinical condition with limitations and difficulties related to immobilization. Human Pose Estimation (HPE) models can be useful as an alternative anthropometric method that allows measuring individual body segments. This study aimed to compare the performance of HPE models for assessing limb length. A physiotherapist evaluated the limb lengths of arms, forearm, trunk, and legs for fifty-nine participants using both tape and five HPE models: PoseNet and four variations of MoveNet. Participants were positioned standing in front of a backdrop while photos were captured using the NLMeasurer mobile application. NLMeasurer is a smartphone application able to record postural, goniometric, and limb length assessments based on the identification of Key Joint Points (KJPs) performed by HPE models. The bias and error measures were employed to evaluate the difference and agreement between measurement methods. Results indicated that the performance of the models was similar, with an average error of 3.5 cm across most segments. However, the models' performance declined for hip-knee distance measurements, with errors ranging from 7.34 to 13.57 cm. HPEs showed promise for facilitating limb length assessments in an automated way, but they have not proven to be accurate enough to be incorporated into clinical practice and used without any human intervention.

ABSTRACT. With the increasing aging population, chronic pain and musculoskeletal disorders have become significant health concerns, leading to a growing demand for non-invasive pain management strategies such as transcutaneous electrical nerve stimulation (TENS). While TENS is widely used for pain relief and neuromuscular regulation, gender differences in physiological responses remain underexplored. This study investigates gender-specific autonomic nervous system (ANS) and electromyography (EMG) responses to TENS stimulation at varying intensities (3VPP, 7VPP, and 11VPP) in 31 participants (16 males, 15 females). ECG and EMG signals were recorded using the BIOPAC MP36 system, and key HRV and EMG parameters were analyzed to assess neural and muscular activation patterns. Results showed that females exhibited broader neuromuscular responses at lower intensities, whereas males demonstrated increased autonomic activity at higher intensities, supporting the Gate Control Theory. These findings highlight the necessity of gender-specific electrostimulation therapy strategies and contribute to optimizing personalized TENS treatments.

ABSTRACT. Accessing reliable clinical knowledge quickly is an everyday challenge for clinicians. Large Language Models (LLMs) can assist healthcare professionals by providing this knowledge, but their responses often deviate from expert consensus or are not up to date necessitating reliable validation and possible correction. To address this, we introduce MedBlock-Bot, an interactive Streamlit-based system integrating a blockchain-enabled Retrieval-Augmented Generation (RAG) framework for expert-driven assessment and immutable feedback storage within a permissioned consortium network. Unlike traditional feedback mechanisms that may be altered or lost, MedBlock-Bot employs smart contracts to securely store and verify any feedback, ensuring transparency and auditability. We evaluated the system using three open-source LLMs—BioMistral, HippoMistral, and LLaMa 3.1—on clinical guideline interpretation for neonates with hypoplastic left heart syndrome. Human experts assessed model responses based on accuracy and relevance, revealing variations in adherence to the guideline knowledge. Additionally, deploying the blockchain component in a local permissioned environment (Ganache) ensured efficient transaction processing and tamper-proof feedback retrieval without gas cost concerns. Our results demonstrate the integration of blockchain for LLM feedback review enhancing trust, accountability, and structured knowledge retention. Clinicians can access past expert assessments for validation, while developers can leverage this feedback for potential model refinement. Taking the long-term impact into account this approach targets towards a reliable and dynamic representation of clinical knowledge and consensus. Open-Source Code: https://github.com/yaseen28/MedBlock-Bot

ABSTRACT. This work proposes an approach based on Large Language Models (LLMs) for creating digital collections from free-text medical reports. The approach uses instruct LLMs to extract relevant clinical terms from these reports, as well as to catalog them using the extracted terms. The cataloged reports are then integrated into a digital collection management platform, enabling further curation by clinical experts. To confirm the feasibility of the approach, we used various models associated with DeepSeek, as well as a coding model developed by Alibaba, and the experimental Clavy reconfigurable collection management platform to handle the resulting collections. The preliminary evaluation results demonstrate the feasibility of the approach, even without relying on external services that could compromise data privacy in real-world scenarios, or on particularly expensive dedicated hardware.

ABSTRACT. Accurately detecting pain in infants remains a complex challenge. Conventional deep neural networks used for analyzing infant cry sounds typically demand large labeled datasets, substantial computational power, and often lack interpretability. In this work, we introduce a novel approach that leverages OpenAI’s vision-language model, GPT-4(V), combined with mel spectrogram-based representations of infant cries through prompting. This prompting strategy significantly reduces the dependence on large training datasets while enhancing transparency and interpretability. Using the USF-MNPAD-II dataset, our method achieves an accuracy of 83.33% with only 16 training samples, in contrast to the 4,914 samples required in the baseline model. To our knowledge, this represents the first application of few-shot prompting with vision-language models such as GPT-4o for infant pain classification.

ABSTRACT. Understanding surgical scenes can provide better healthcare quality for patients, especially with the vast amount of video data that is generated during MIS. Processing these videos generates valuable assets for training sophisticated models. In this paper, we introduce CLIP-RL, a novel contrastive language-image pre-training model tailored for semantic segmentation for surgical scenes. CLIP-RL presents a new segmentation approach which involves reinforcement learning and curriculum learning , enabling continuous refinement of the segmentation masks during the full training pipeline. Our model has shown robust performance in different optical settings, such as occlusions, texture variations, and dynamic lighting, presenting significant challenges. CLIP model serves as a powerful feature extractor, capturing rich semantic context that enhances the distinction between instruments and tissues. The RL module plays a pivotal role in dynamically refining predictions through iterative action-space adjustments. We evaluated our model CLIP-RL on the EndoVis2018 and EndoVis2017 datasets. CLIP-RL achieved a mean IoU of 81\% , outperforming state-of-the-art models and a mean IoU of 74.12\% on Endovis 2017. This superior performance was achieved due to the combination of contrastive learning with reinforcement learning and curriculum learning.

ABSTRACT. Effective visualization of three-dimensional (3D) medical images is essential for planning medical procedures, as it directly impacts diagnostic precision and enhances patient understanding. High-quality 3D visualization enables physicians to make informed decisions, supports surgeons in planning complex procedures, and helps patients better understand their conditions and treatment options. However, commonly used head-mounted display (HMD) devices such as the Meta Quest and Microsoft HoloLens often cause discomfort, eye strain, and communication challenges, and their complex setup can limit practical use in clinical settings. The recently developed, glasses-free Looking Glass Factory 16'' Spatial Display presents a compelling alternative, capable of delivering high-quality holographic images without requiring HMDs, making it a promising solution for clinical visualization.

This study investigates using the Looking Glass 16'' Spatial Display to visualize 3D pediatric cardiac CT scans in DICOM format, demonstrating its potential to improve accessibility and usability in clinical imaging. Clinical DICOM images are converted into MetaImages (RAW images with MHD headers) using 3DSlicer software and rendered with a volume rendering algorithm to create detailed volumetric models. These 3D images are then displayed on the Looking Glass device through the Unity3D platform, with an NVIDIA GeForce RTX 4060 Ti graphics card optimizing frame rates, reducing latency, and supporting real-time image control on a personal computer (PC). Experimental evaluations confirm the feasibility of producing high-quality, real-time displays on the Looking Glass 16'' Spatial Display, offering clinicians an intuitive and efficient interface. This research highlights the potential of the Looking Glass 16'' Spatial Display to enhance 3D medical image visualization, particularly for cardiac CT applications. By providing a more precise, accessible imaging method, this technology could significantly improve clinical decision-making and deepen medical professionals’ understanding of complex anatomical data.

ABSTRACT. Lung cancer has the highest mortality rate among all cancer types, affecting both men and women. It is estimated that lung cancer accounts for 21% of cancer deaths in each gender worldwide. In Brazil, lung cancer is the third most common type of cancer among men and the fourth most common among women. This alarming statistic highlights the significant impact of lung cancer on overall cancer mortality, underscoring the urgent need for effective prevention, early detection, and treatment strategies to combat this disease. The Lung-RADS is a standardized classification system for lung nodules detected in imaging exams and assesses the risk of malignancy (cancer) in these nodules to guide subsequent management decisions. In this context, the main goal of this work was to evaluate the effectiveness of using question-answering natural language processing using LLMs to extract lung nodule characteristics from Portuguese chest CT reports for automated Lung-RADS classification. Our most effective model was Llama 3.3 70B, and this model achieved a weighted F1 score of 84.77%. Our findings underscore the potential of LLMs to support radiologists in accurately categorizing lung nodules according to Lung-RADS criteria, thereby simplifying the diagnostic process.

ABSTRACT. Cone Beam Computed tomography (CT) imaging is widely used in diverse clinical applications, but its field of view (FOV) is often limited by the detector size, leading to truncation artifacts that obscure anatomical structures and compromise diagnostic accuracy. While several deep learning solutions have demonstrated success in extending the FOV, they primarily rely on spatial inpainting and do not fully exploit the frequency-domain characteristics of sinograms. Besides, some of the most successful solutions are slow as they work both in the projection and reconstruction domain or use heavy networks, hindering their implementation in real clinical practice. In this work we propose a recurrent network that works in the sinogram domain solely and that incorporates Fast Fourier Convolution blocks at early feature extraction stages, to enable larger receptive fields and enhance sinogram extrapolation by leveraging spectral domain information. Our approach aims to achieve accurate and efficient sinogram completion, reducing truncation artifacts while maintaining real-time clinical feasibility. The proposed method holds potential for improving CT image reconstruction quality, enhancing diagnostic accuracy, and expanding the applicability of C-arm CT systems in various medical fields.

ABSTRACT. In Cone Beam CT (CBCT) systems, each acquisition is carried out with a stationary bed. Therefore, the field of view is limited by the size of the detector. To overcome this limitation, it is common to perform successive acquisitions for different bed positions, which are subsequently combined to increase the field of view in the longitudinal direction. To avoid the appearance of double edges in the resulting volume, it is necessary to calculate the exact bed displacement, which has traditionally been obtained by prior geometric calibrations with calibration phantoms. This implies that the calibration must be repeated periodically to adapt to any changes that the equipment may undergo. As an alternative, in this work we propose the use of an automatic calibration algorithm capable of obtaining the misalignment parameters in real time, allowing the obtaining of multi-bed images in small and regular animal CBCT systems without the need of a previous calibration.

ABSTRACT. Urban vulnerability assessment is crucial for understanding the spatial distribution of deprived areas and associated risks. Slum residents face significantly worse health outcomes than non-slum urban populations, with neighborhood effects being critical in social epidemiology. Identifying such areas is vital because they present public health challenges that climate change and increased air pollution can exacerbate. Accordingly, this study proposed an AI-driven methodology that integrates remote sensing data, socioeconomic indicators, and machine learning algorithms to identify and analyze vulnerable areas in Brazil. To create a vulnerability index, we incorporate multiple data sources, including Sentinel-2 and Sentinel-5P imagery, Brazilian socioeconomic indicators, and OpenStreetMap. Hence, we predicted pollution indicators using regression algorithms such as Random Forest, XGBoost, and Linear Regression. Our findings demonstrate that integrating multi-source data is a promising approach for better understanding deprived areas, indicating that slums (called "favelas" in Brazil) exhibit an intense concentration of the socioeconomic vulnerability index, a key determinant of deprivation. However, non-slum areas may present heterogeneous conditions, with some regions showing vulnerability levels comparable to those of slums while others show better conditions. Our results highlight the potential of AI-driven approaches for urban vulnerability assessment, offering insights for policymakers and researchers.

ABSTRACT. Detection of fraud, waste and abuse in healthcare systems is critical to minimise financial losses incurred by medical aid companies. Traditional fraud detection methods are based on manual processes and often fail to identify hidden patterns within medical claims data. The complexity and scale of fraudulent activities require more advanced, data-driven approaches. This study aims to develop a framework that employs unsupervised machine learning models to aid in and improve fraud detection from medical claims data for a South African medical scheme. The framework employs six unsupervised anomaly detection algorithms, including isolation forest (IF), one-class support vector machine (OC-SVM), autoencoder (AE), self-organising map (SOM), local outlier factor (LOF), and hierarchical density-based spatial clustering of applications with noise (HDBSCAN). The first layer is focused on detecting global outliers, and the second layer is aimed at refining those classifications with locally outlier factors. The results show that the framework flags potential fraud, waste, and abuse and provides valuable insights into which medical practices should be investigated.

ABSTRACT. This study examines the current state and impact of serious games in Ecuador's healthcare sector, focusing on their applications in medical education, rehabilitation, and behavioral change. Through a systematic review of 30 articles (2018–2024), the research highlights successful implementations, such as "ITaCaS" for sexual health education, "FarmFunTime" and "PICTOAPRENDE" for physical and cognitive rehabilitation, and "Escape Room for Healthy Habits" for promoting lifestyle changes. Despite these advancements, challenges like limited funding, cultural barriers, and lack of standardized protocols hinder widespread adoption. The study identifies critical gaps in maternal health, chronic disease management, and elderly care, urging future efforts to prioritize low-resource adaptations, interdisciplinary collaboration, and longitudinal studies. The findings underscore the transformative potential of serious games in Ecuadorian healthcare while emphasizing the need for culturally tailored and scalable solutions to address systemic inequities.

ABSTRACT. Health-related quality of life (HRQoL) is a multidimensional construct reflecting individuals’ overall well-being, including physical, mental, emotional, and social aspects. This study focuses on assessing HRQoL among university students. A comprehensive survey, incorporating several validated instruments was administered to a stratified sample of undergraduate and graduate students. To analyze the collected data, we developed two deep learning models: a Fully Connected Neural Network (FCNN) and a Graph Neural Network (GNN). The GNN model represents each student’s responses as a tree-like graph and processes this structured data through multiple convolutional layers enhanced by a sort-pooling layer, which standardizes graphs of varying dimensions. Experimental results indicate that both models effectively classify students based on their HRQoL profiles. Notably, the GNN model achieves higher test accuracy and F-Score compared to the FCNN, demonstrating robust performance even when analyzing incomplete surveys. These findings underscore the potential of graph-based deep learning methods by integrating heterogeneous data and capturing complex relationships, these approaches offer a promising solution for monitoring and improving students well-being.

ABSTRACT. Magnetic Resonance Imaging (MRI) plays an important role in the diagnosis and monitoring of central nervous system (CNS) pathologies. There are various MRI modalities that enable the visualization of distinct structural features in neoplasms. Among these modalities, contrast-enhanced weighted T1 images (T1W-CE) are widely utilized, as they clearly delineate tumor boundaries and highlight cerebral vasculature. This modality relies on gadolinium-based contrast agents (GBCAs). However, due to risk of deposition in tissues after repeated administration and expensiveness, it is recommended to limit their use. To address this limitation, this study proposes a deep learning (DL) algorithm that generates synthetic T1W-CE images from non-contrast MRI sequences. The preliminary evaluation was performed using conventional, quantitative metrics and qualitative assessment by an expert neuroradiologist. This evaluation supports the potential feasibility of the proposed method in reconstructing synthetic T1W-CE MRI scans that closely resemble real contrast-enhanced images.

ABSTRACT. Triple-negative breast cancer (TNBC) is an aggressive subtype with limited treatment options and a poor prognosis,necessitating accurate early diagnosis to optimize therapeutic interventions. This study aims to develop a predictive method using MRI radiomics and machine learning to distinguish TNBC from other breast cancer subtypes. MRI data from 87 patients with invasive breast cancer were retrospectively analyzed. Manual segmentation of dynamic contrast-enhanced MRI (DCE) was performed, and the segmented masks were propagated to T1- weighted, T2-weighted water, and fat scans. Radiomic features were extracted using PyRadiomics, and feature selection was performed using Spearman’s correlation, mutual information, and least absolute shrinkage and selection operator (LASSO). The EasyEnsemble classifier, an ensemble of AdaBoost learners trained on balanced bootstrap samples, was employed for classification. The combination of DCE and T1W and T2F MRI modalities consistently outperformed individual modalities. LASSO feature selection resulted in the most significant performance improvements, with the highest area under the curve (AUC-score) of 0.93 ± 0.05, balanced accuracy of 0.81 ± 0.04, and F-score of 0.74 ± 0.05. These findings demonstrate the potential of MRI radiomics and machine learning to noninvasively enhance the diagnostic capability of TNBC, thereby contributing to improved patient care and personalized treatment strategies.

ABSTRACT. Early detection of Alzheimer’s disease is critical for timely intervention, and neuroimaging biomarkers play a fundamental role in assessing structural brain changes. The Koedam visual scale is a widely used tool for evaluating parietal atrophy, particularly in early-onset AD. This study presents an automated approach to Koedam scale classification using T1-weighted MRI features and clustering techniques. The proposed method follows a structured pipeline, including skull stripping, noise reduction, bias field correction, and region of interest (ROI) selection. Brain tissue segmentation is performed using a probabilistic model-based approach, classifying image voxels into gray matter, white matter, and cerebrospinal fluid. Additionally, deformation fields derived from nonlinear image registration with a non-atrophied template are extracted to capture structural differences associated with atrophy. The strain tensor, derived from the displacement field, is computed to further characterize tissue deformation. A feature selection step is applied before clustering, where a Gaussian Mixture Model (GMM) clustering algorithm is used to categorize images into four Koedam atrophy levels, mimicking expert visual assessment. The method was evaluated on a dataset of 103 MRI images, demonstrating a clear differentiation between atrophy severity levels. The resulting clusters exhibited progressively increasing mean Root Mean Square displacement magnitude (RMSdm) values: 0.56±0.08 for cluster 0, 0.59±0.06 for cluster 1, 0.84±0.07 for cluster 2, and 1.09±0.14 for cluster 3. These findings indicate that the proposed approach effectively quantifies parietal atrophy, providing an objective and reproducible alternative to expert visual assessment.

ABSTRACT. Magnetic resonance imaging (MRI) is invaluable for the detailed visualization of soft tissues. However, its susceptibility to motion artifacts presents a challenge in ophthalmology due to the continuous eye movement. A recently developed technique effectively resolves eye motion in MRI, but it strongly relies on an eye tracker (ET), which, being a resource-intensive system, limits its broader adoption. The present work introduces a novel approach that, based on machine learning techniques, enables automated identification of eye motion in raw MRI data collected using a fast-sampling acquisition strategy. Such MRI data were acquired from nine healthy subjects while visual stimuli directed their gaze. A synchronized ET signal was also recorded to label the data. A broad spectrum of features representing the raw MRI acquisitions were extracted, and the classification models were built based on the most discriminative ones. Motion identification was primarily driven by phase-related features, while the models achieved accuracy and recall values exceeding 98%. This study represents an important first step towards obtaining a high-quality MRI of the eye without depending on supplementary hardware.

ABSTRACT. This study presents the first phase of a transdisciplinary research project aimed at improving the design of visual feedback stimuli in neurofeedback (NFB) applications. While current NFB research has focused extensively on signal processing and feature extraction, limited attention has been given to the design and user experience of feedback stimuli. To address this gap, the research team conducted generative user research including site visits, expert consultations, and semi-structured interviews with domain experts and previous NFB participants. Thematic analysis of the collected data yielded a preliminary set of design requirements.

User-centered requirements include minimizing cognitive load, enhancing attention and engagement, incorporating positive reinforcement, supporting a sense of agency, and providing clear instructions. Technical requirements include reducing artifacts, ensuring low-latency feedback, and promoting participant relaxation.

These findings lay the groundwork for iterative design and evaluation phases, with the ultimate goal of delivering validated stimuli and design guidelines to the NFB research and clinical communities.

ABSTRACT. Tinnitus (TI) is a disease of the brain with high prevalence and often severe consequences for which no causal therapy approach has been established so far. Neurofeedback Training (NFT) is considered a promising approach to treat TI based on studies applying the Dohrmann-protocol reporting reduced TI loudness and distress. As the method is relatively laborious and expensive, home-based NFT could make this promising method accessible to a larger number of patients. However, it is still unclear whether and how NFT can be carried out at home. This study evaluated the feasibility of the Dohrmann-protocol in a home-based, sham-controlled, single blind, longitudinal cross-over wash-out design with N = 9 TI patients. EEG was recorded during 30 NFT or sham feedback sessions and acceptance of the at-home treatment was measure longitudinally. Ordinary acceptance, especially in response to veritable NFT in comparison to sham feedback and a drop-out rate of 22.20% were observed. All home-based NFT sessions produced impedances <10 kOhm. TI distress was reduced, and NFT increased the alpha delta ratio. We conclude, the feasibility of a methodologically sound home-based NFT study was demonstrated. Limitations discuss the small sample size.

ABSTRACT. Eye-tracking technology has been increasingly utilized in various medical domains, yet its adoption in pathology remains limited. The lack of standardized methodologies and the complexity of analyzing whole-slide images pose significant challenges for applying eye-tracking in this domain. Existing studies often rely on proprietary and heterogeneous software solutions, reducing reproducibility and comparability of results across researchers. To address these issues, we present a novel eye-tracking study platform specifically designed for digital pathology. The proposed platform enables standardized and reproducible eye-tracking studies by providing a modular architecture that supports common eye-tracking software and hardware. It features a web-based frontend for study execution and a backend deployed via Docker, ensuring platform-independent usability while maintaining local data storage for privacy compliance and GDPR adherence. An experimental study was conducted with pathologists, trainees, and medical students to validate the applicability of the platform, as well as to demonstrate its reliable performance and ease-of-use. While the platform proved technically robust, areas for improvement remain. Future enhancements will focus on refining the user experience by integrating an improved tutorial system and post-task feedback mechanisms, incorporating gamification elements to boost participant engagement and data quality. Additionally, the next major version will introduce full support for whole-slide images, including zooming and advanced navigation features, to provide more comprehensive insights into pathologists’ visual attention patterns. By providing a structured and adaptable research framework, our work represents a significant step toward standardizing eye-tracking research in pathology. The developed platform provides the foundation for more consistent study designs and reproducible findings, ultimately contributing to the advancement of digital pathology and diagnostic training methodologies.

ABSTRACT. Eye tracking has shown promise in detecting group-wise differences between healthy controls and people with Mild Cognitive Impairment (MCI) who might progress to develop Alzheimer's disease. As there is currently no cure, only medications to slow the progress of the disease, it is of paramount importance to find the persons at risk early.

We extracted saccade- and Region of Interest (ROI)-centric feature sets from a clinical eye tracking dataset and analyzed the results using machine learning to establish which features were most beneficial to correctly classify individuals at risk.

Our results show that the analysis of multimodal eye tracking recordings of number and text reading tasks produces feature sets that can have good predictive value to classify MCIs from healthy controls. Changes in participants' pupil sizes from the personalized baseline appear to be especially promising candidates for improving the classification efficiency of MCIs.

ABSTRACT. Cognitive impairment is a growing public health concern, with early detection playing a crucial role in improving patient outcomes. The Montreal Cognitive Assessment (MoCA) is widely used for screening mild cognitive impairment (MCI) and early-stage dementia. However, traditional MoCA assessments require manual scoring by trained professionals, making the process labor-intensive, time-consuming, and susceptible to human error. To overcome these limitations, we propose an automated pipeline for MoCA score estimation using eye-gaze data and Vision Transformers (ViTs). Our approach leverages gaze-tracking technology to capture spatial and temporal eye-movement patterns during structured cognitive tasks, identifying subtle cognitive impairments that may otherwise go unnoticed. The raw gaze data is preprocessed and mapped onto task-relevant image regions, where a pretrained ViT extracts high-dimensional feature representations. To address inconsistencies in gaze sampling and improve temporal modeling, we introduce a time-aware positional embedding mechanism that enhances the model’s ability to infer cognitive performance. These extracted features are then processed by a transformer-based classification model to predict MoCA scores with high accuracy. We validate our approach using a dataset collected from seven cognitive gaming sessions, demonstrating its effectiveness in automated cognitive assessment. The experimental results indicate that our method provides a reliable and efficient alternative to traditional MoCA evaluations, reducing dependency on human intervention while maintaining diagnostic accuracy.

ABSTRACT. Retinopathy is characterized by pathological alterations to the retina that can lead to partial or total vision loss. It can result from a variety of causes, including diabetes,hypertension, and autoimmune disorders, such as Autoimmune Retinopathy, which is particularly challenging to detect and diagnose. Due to its visually similar features, Autoimmune Retinopathy is often mistaken for other retinal diseases that require distinct treatments. The objective of this work is to develop an anomaly detection system for retinal images using Transformer-based deep learning models. By leveraging Transformers, the model captures subtlevariations in retinal images, enhancing diagnostic accuracy. This approach aims to assist ophthalmologists in precise disease detection, improving early diagnosis and enabling personalized treatment strategies.

ABSTRACT. Obstructive Sleep Apnea (OSA) management struggles from static retrospective adherence models which build their assessment on rigid compliance benchmarks. The study introduces an Artificial Intelligence-based framework for CPAP treatment adaptation which combines real-time biosensor data and personalized treatment stages. The system categorized patients based on therapy dynamics through unsupervised learning analysis into four stages new, adherent, attempting, and non-adherent while considering CPAP usage data and clinical histories and Monitoring survey results. The adopted monitoring system tracks a selection of digital biomarkers (e.g., HRV, ODI, RVO) in real-time to identify initial signs of therapy deviation as well as changes in physiological responses. It enables the risk scoring system to set off weekly risk evaluations and adjust cluster groupings and activate custom-made intervention plans such as digital prompts and medical survey requests alongside physician alert actions. Continuous physician feedback connections enable healthcare providers to check and modify or restructure interventions to ensure clinical accuracy and security. The system provides dynamic adjustment of biomarker approaches alongside intervention force levels which depend on patients' current therapy state and clinical complexity. The proposed framework displays promising results to detect patients at risk for dropping out and to optimize their therapy according to preliminary retrospective data analysis. The developed work provides a scalable and explainable clinical framework that will support future tests and deployments for biomarker-driven real-time CPAP personalization methods throughout the OSA care pathway.

ABSTRACT. Artificial intelligence (AI) is making significant advances toward becoming a well-established and promise-bearing technology in various medical domains such as screening, diagnostics, and biopharma research. However, its state remains relatively nascent in surgery and surgical therapeutics. This presents an opportunity for leveraging ongoing rapid advances in AI technology and the increasing availability of large, diverse datasets to pave the way for their use in these domains. Expanding the use of AI to include various processes in surgery-related workflows could provide several benefits, such as greater assurance for reduced errors, better assistance to surgeons, and overall improved patient outcomes. To encourage further research in surgical AI, this article summarizes the state-of-the-art in AI assurance in various aspects of a patient's timeline when undergoing head and neck surgeries, including diagnostics, preoperative considerations, intraoperative guidance, and postoperative and outcome predictions. The work aims to highlight gaps in the state-of-the-art and identify opportunities for the computer-based medical systems community to encourage future research and development on the subject.

ABSTRACT. Plane wave imaging (PWI) has significantly ad vanced ultrasound diagnostics by enabling high frame rates and real-time capabilities critical for applications such as cardiac imaging and elastography. However, its clinical utility remains constrained by inherent trade-offs between spatial resolution, contrast, and frame rate, particularly when using a limited number of transmissions. Existing solutions, such as pixel-level apodization or angle-dependent transmit apodization (ADTA), either impose prohibitive computational costs or restrict adapt ability to small fields of view (FOV). To address these limitations, we propose a genetic algorithm (GA)-based framework for depth-dependent transmit apodization optimization in PWI. Our approach models the spatial contribution of each plane-wave transmission via a continuous parametric function, reducing the optimization dimensionality. The multi-objective GA simulta neously minimizes the full width at half maximum (FWHM) and peak sidelobe level (PSL), achieving resolution enhancement and artifact suppression across the entire FOV. Experimental validation using a Verasonics Vantage 256 system and an L7 4 linear probe on a tissue-mimicking phantom demonstrate the feasibility of the approach using only seven steered transmissions.

ABSTRACT. Divergent wave imaging (DWI) achieves high-frame rate ultrasound with a broad field of view incomparable to conventional line-by-line ultrasound techniques. However, unlike traditional focused ultrasound, which applies apodization in both transmit and receive modes, DWI restricts beamforming adjustments to the receive phase, inherently limiting its ability to suppress off-axis clutter and noise. To address this, we propose a multi-objective optimization framework using genetic algorithms to derive spatial weights for DWI transmissions. Each virtual source’s contribution is modeled as a depth-dependent Gaussian function, with optimized parameters—lateral spread, axial coverage, and amplitude—to enhance resolution and con trast. The approach was validated through in silico data using convex probe, demonstrating a 25% reduction in full-width-at half-maximum (FWHM) and a 44% improvement in contrast ratio (CR) compared to conventional compounding with only ten virtual sources.