Abstract
Diabetic retinopathy (DR) is a vision-threatening complication of diabetes and one of the leading causes of blindness worldwide. It is characterized by the appearance of lesions on the retina, such as microaneurysms, hemorrhages, hard exudates, and soft exudates, which are crucial for staging the disease. Diagnosis is typically performed through analysis of fundus images, a manual process that is time-consuming and prone to subjectivity. To address this, this study explores the automatic segmentation of DRrelated lesions using deep learning techniques. Four convolutional neural network architectures were evaluated: U-Net, FPN, DeepLabV3+, and Attention U-Net. The IDRiD dataset was used for training and validation The DeepLabV3+ model with ResNet50 achieved the highest overall performance, while FPN was the only model capable of detecting microaneurysms in the multiclass task. These findings underscore the importance of architecture selection, loss function design, and preprocessing choices. Future work may explore new datasets, enhanced data augmentation, and the impact of optic disc removal on segmentation accuracy.

Abstract
Retinal diseases such as glaucoma, diabetic retinopathy, and age-related macular degeneration affect hundreds of millions of people worldwide and are among the leading causes of vision loss. Optical Coherence Tomography (OCT) is a non-invasive imaging technique widely used to support the diagnosis of these conditions. However, manual analysis of OCT images is time-consuming, prone to inter-observer variability, and requires extensive clinical expertise. In recent years, deep learning methods have shown outstanding performance in medical image segmentation tasks. This work proposes an automatic approach for the segmentation of retinal layers in OCT images using the GOALS 2022 dataset. Four segmentation architectures were evaluated - U-Net, DeepLabV3+, FPN (U-Net++), and Attention U-Net - all combined with the ResNet50 encoder. Additionally, the influence of encoder selection in the U-Net architecture was investigated, testing ResNet34, EfficientNetB0, MobileNetV2, VGG16, and InceptionV3. The results show that the DeepLabV3+ model achieved the best overall performance, with an F1-Score of 0.9669 and an IoU of 0.9370. These findings demonstrate that lightweight, accessible models can achieve results comparable to state-of-the-art methods, offering a promising solution for clinical applications in retinal image segmentation. © 2025 The Authors. Published by Elsevier B.V.

Abstract
High-resolution UAV imagery offers unprecedented opportunities for vineyard monitoring, yet its practical use in semantic segmentation is hindered by the high cost of pixel-level annotation. Weakly supervised learning (WSL) emerges as a promising alternative, capable of reducing annotation effort while preserving competitive performance. In this study, we conduct a direct comparative evaluation of two pseudo-labelling strategies for vine row segmentation, a task still underexplored in perennial crops. The first strategy combines a spectral heuristic with Conditional Random Fields (CRF) to enforce spatial consistency, while the second employs token clustering of DINO-ViT embeddings. To ensure fairness, both pseudo-label sets were used to train an identical segmentation architecture (U-Net with ResNet50), thereby isolating the impact of pseudo-label quality. Results, measured by precision, recall, F1-score, and Intersection over Union (IoU), reveal that the CRF-refined heuristic (F1 = 0.77, IoU = 0.62) consistently outperforms the transformer-based clustering approach (F1 = 0.52, IoU = 0.50). These findings highlight the decisive role of spatial regularisation in weak supervision and provide a reproducible pipeline that balances accuracy, methodological simplicity, and annotation cost. The contribution of this work lies in demonstrating a practical and extensible framework for UAV-based vineyard monitoring, while opening pathways for hybrid approaches that integrate semantic depth with spatial coherence in future research.

Abstract
Video capsule endoscopy (VCE) enables high-resolution visualisation of the small bowel but remains constrained by manual review of thousands of frames, which is time-consuming and error-prone under class imbalance. This study investigates deep learning for automatic multiclass lesion classification in VCE, comparing two convolutional networks (ResNet-50, EfficientNet-B3) with two Vision Transformers (Swin, DeiT) on the public Kvasir-Capsule dataset (47,161 images; 11 classes). The pipeline comprises standard preprocessing, class-aware augmentation and adaptive data augmentation, stratified data partitioning, hyperparameter optimisation with Optuna, and evaluation using accuracy, precision, recall, and F1-score. DeiT achieved the best overall performance (accuracy = 0.98; F1 = 0.96), with strong class-wise results in clinically salient categories (e.g., ulcer, fresh blood, angiectasia), indicating effective modelling of long-range dependencies and subtle patterns. We further assess computational feasibility by reporting training configuration and indicative inference time per image, supporting potential integration into assisted reading workflows. Limitations include reliance on a single public dataset, pronounced class imbalance, and the absence of prospective clinical validation, which may affect generalisability. These findings position Transformer-based models as promising candidates for VCE decision support, while underscoring the need for future work on (i) multicentric datasets and external validation, (ii) comprehensive statistical analysis with confidence intervals and robust baselines under imbalance, and (iii) prospective studies quantifying end-to-end impact on reading time and diagnostic safety. © 2025 The Authors. Published by Elsevier B.V.

Abstract
Video Capsule Endoscopy (VCE) is a pivotal technology in modern gastroenterology, offering a non-invasive method to visualize the entire small bowel. However, the clinical application of VCE is hampered by the extensive review time required, as specialists must manually analyze thousands of images from each procedure. This process is not only laborious and costly but also prone to diagnostic errors due to fatigue, subtle abnormalities, and variability in interpretation across clinicians. To address this challenge, deep learning methods have been explored to automate VCE image analysis. However, most existing approaches rely on a single model architecture, which often fails to generalize across the broad visual diversity found in gastrointestinal imagery. This limitation becomes especially pronounced in multiclass classification tasks, where the ability to distinguish between visually similar tissues and lesions is essential. Ensemble-based methods such as Mixture of Experts (MoE) have shown promising results in general computer vision by leveraging multiple specialized models for improved robustness. However, no prior work has investigated MoE or Hierarchical MoE (HMoE) architectures for multiclass classification of VCE or endoscopic images more broadly. To explore this opportunity, we present a comparative framework evaluating three deep learning strategies for VCE image classification: individual models, flat MoE systems, and Hierarchical MoE architectures. Using a subset of the Kvasir-Capsule dataset, which contains 12 gastrointestinal tissue and lesion classes, we first train and evaluate four backbone models (InceptionNeXt, EfficientViT, ConvNeXtV2, and DeiT3) to establish a performance baseline. The two best-performing architectures, ConvNeXtV2 and DeiT3, are then used as expert backbones within both MoE and HMoE systems. In the MoE configuration, a gating network assigns dynamic per-image weights to multiple expert instances. In contrast, the HMoE configuration constructs a learned binary tree that routes samples based on class similarity through increasingly specialized branches. In the HMoE models, ConvNeXtV2 outperformed DeiT3 in accuracy, whereas DeiT3 showed superior routing accuracy. These results indicate that expert-driven ensemble methods not only outperform standalone models but also offer complementary advantages depending on architecture and routing strategy. This study provides new evidence for the clinical potential of MoE and HMoE frameworks in scalable, accurate VCE image analysis.

Abstract
Glaucoma is one of the leading causes of irreversible blindness worldwide, affecting millions of people, often silently and progressively. Early diagnosis is crucial to slow its progression, but it remains challenging due to the need for manual analysis of large volumes of retinal images by trained specialists. In this context, automatic detection systems based on deep learning offer a promising opportunity to facilitate and accelerate the diagnostic process, providing scalability and high accuracy. This work presents the development of an automatic method for optic disc and optic cup segmentation in retinal fundus photographs, aiming to support early glaucoma detection. The proposed methodology is based on convolutional neural networks (CNNs), specifically an enhanced U-Net architecture with a ResNet50 backbone, incorporating attention mechanisms and data augmentation strategies to improve segmentation accuracy. The model was trained and validated using the REFUGE dataset, which contains high-quality fundus images with manual annotations of the disc and cup regions. Experimental results demonstrate that the developed model achieved an average Dice coefficient of 0.937 for optic disc segmentation and 0.828 for optic cup segmentation. Analysis of the cup-to-disc ratio (CDR) yielded mean values of VCDR = 0.497 ± 0.059, ACDR = 0.252 ± 0.060, and mean CDR = 0.375 ± 0.058, with 55.0% of cases classified as low risk, 43.3% as moderate risk, and 1.7% as high risk for glaucoma. These results highlight the potential of the proposed method as an assistive tool for automated glaucoma screening.