DEVELOPMENT AND EVALUATION OF AN EXPLAINABLE 3D CONVOLUTIONAL NEURAL NETWORK FOR HEALTHY–PATHOLOGICAL CLASSIFICATION OF RETINAL OCT VOLUMES

Optical Coherence Tomography (OCT) has become indispensable for the diagnosis of retinal diseases. However, the manual interpretation of volumetric data is time-consuming and prone to inter-observer variability. While Deep Learning (DL) models has demonstrated significant potential in automating this process, current literature predominantly relies on two-dimensional models that neglect the spatial continuity of the retinal stack. Furthermore, the ``black-box'' nature of these algorithms hinders clinical trust and accountability. This thesis addresses these limitations by developing and evaluating an explainable 3D Convolutional Neural Network (3D CNN) framework for the binary classification of healthy and pathological OCT volumes. Utilizing a large-scale standardized dataset of 4,942 volumetric acquisitions, we implemented a 3D ResNet-18 architecture. Experimental results established that pre-training on the Kinetics-400 domain significantly outperforms medical cross-domain initializations, enabling the model to effectively capture volumetric biomarkers. To ensure clinical transparency, the system integrates a multimodal explainability framework combining Grad-CAM and perturbation-based methods such as Occlusion and LIME. This is further enhanced by LO-VLM, a visual language model which translates visual saliency into structured semantic descriptions. Quantitative evaluation demonstrates the superiority of the volumetric approach, achieving an AUC of 0.862 and a specificity over 90%. The model achieved a significant reduction in false negatives subjects, a vital requirement for patient safety in screening scenarios. Qualitative analysis through representative case studies confirmed that the visual-semantic consensus between saliency maps and LO-VLM descriptions effectively resolves decisional uncertainty in borderline cases. Ultimately, this work provides a robust, traceable, and clinician-friendly tool that bridges the gap between high-performance Artificial Intelligence (AI) and the transparency requirements of modern ophthalmological practice.