Abstract

Abstract
Medical case retrieval plays a crucial role in clinical decision-making by enabling healthcare professionals to find relevant cases based on patient records, diagnostic images, and textual descriptions. Given the inherently multimodal nature of medical data, effective retrieval requires models that can bridge the gap between different modalities. Traditional retrieval approaches often rely on unimodal representations, limiting their ability to capture cross-modal relationships. Recent advances in dense model-based techniques have shown promise in overcoming these limitations by encoding multimodal information into a shared latent space, facilitating retrieval based on semantic similarity. This paper investigates the potential of dense models to enhance multimodal search systems. We evaluate various dense model-based approaches to assess which model characteristics have the greatest impact on retrieval effectiveness, using the medical case-based retrieval task from ImageCLEFmed 2013 as a benchmark. Our findings indicate that different dense model approaches substantially impact retrieval effectiveness, and that applying the CombMAX fusion methodto combine their output results further improves effectiveness. Extending context length, however, yielded mixed results depending on the input data. Additionally, domain-specific models—those trained on medical data—outperformed general models trained on broad, non-specialized datasets within their respective fields. Furthermore, when text is the dominant information source, text-only models surpassed multimodal models

Abstract
In football, where minor differences can significantly affect outcomes and performance, automatic video analysis has become a critical tool for analyzing and optimizing team strategies. However, many existing solutions require expensive and complex hardware comprising multiple cameras, sensors, or GPS devices, limiting accessibility for many clubs, particularly those with limited resources. Using images and video from a moving camera can help a wider audience benefit from video analysis, but it introduces new challenges related to motion. To address this, we explore an alternative homography estimation in moving camera scenarios. Homography plays a crucial role in video analysis, but presents challenges when keypoints are sparse, especially in dynamic environments. Existing techniques predominantly rely on visible keypoints and apply homography transformations on a frame-by-frame basis, often lacking temporal consistency and facing challenges in areas with sparse keypoints. This paper explores the use of estimated motion information for homography computation. Our experimental results reveal that integrating motion data directly into homography estimations leads to reduced errors in keypoint-sparse frames, surpassing state-of-the-art methods, filling a current gap in moving camera scenarios.

Abstract
The large gap between the generalization level of state-of-the-art machine learning and human learning systems calls for the development of artificial intelligence (AI) models that are truly inspired by human cognition. In tasks related to image analysis, searching for pixel-level regularities has reached a power of information extraction still far from what humans capture with image-based observations. This leads to poor generalization when even small shifts occur at the level of the observations. We explore a perspective on this problem that is directed to learning the generative process with causality-related foundations, using models capable of combining symbolic manipulation, probabilistic reasoning, and pattern recognition abilities. We briefly review and explore connections of research from machine learning, cognitive science, and related fields of human behavior to support our perspective for the direction to more robust and human-like artificial learning systems.

Abstract
Traumatic Brain Injury (TBI) is a form of brain injury caused by external forces, resulting in temporary or permanent impairment of brain function. Despite advancements in healthcare, TBI mortality rates can reach 30%–40% in severe cases. This study aims to assist clinical decision-making and enhance patient care for TBI-related complications by employing Artificial Intelligence (AI) methods and data-driven approaches to predict decompensation. This study uses learning models based on sequential data from Electronic Health Records (EHR). Decompensation prediction was performed based on 24-h in-mortality prediction at each hour of the patient's stay in the Intensive Care Unit (ICU). A cohort of 2261 TBI patients was selected from the MIMIC-III dataset based on age and ICD-9 disease codes. Logistic Regressor (LR), Long-short term memory (LSTM), and Transformers architectures were used. Two sets of features were also explored combined with missing data strategies by imputing the normal value, data imbalance techniques with class weights, and oversampling. The best performance results were obtained using LSTMs with the original features with no unbalancing techniques and with the added features and class weight technique, with AUROC scores of 0.918 and 0.929, respectively. For this study, using EHR time series data with LSTM proved viable in predicting patient decompensation, providing a helpful indicator of the need for clinical interventions. © 2025 Elsevier Ltd

Abstract
Lung cancer stands as the most prevalent and deadliest type of cancer, with adenocarcinoma being the most common subtype. Computed Tomography (CT) is widely used for detecting tumours and their phenotype characteristics, for an early and accurate diagnosis that impacts patient outcomes. Machine learning algorithms have already shown the potential to recognize patterns in CT scans to classify the cancer subtype. In this work, two distinct pipelines were employed to perform binary classification between adenocarcinoma and non-adenocarcinoma. Firstly, radiomic features were classified by Random Forest and eXtreme Gradient Boosting classifiers. Next, a deep learning approach, based on a Residual Neural Network and a Transformer-based architecture, was utilised. Both 2D and 3D CT data were initially explored, with the Lung-PET-CT-Dx dataset being employed for training and the NSCLC-Radiomics and NSCLC-Radiogenomics datasets used for external evaluation. Overall, the 3D models outperformed the 2D ones, with the best result being achieved by the Hybrid Vision Transformer, with an AUC of 0.869 and a balanced accuracy of 0.816 on the internal test set. However, a lack of generalization capability was observed across all models, with the performances decreasing on the external test sets, a limitation that should be studied and addressed in future work.