Contribuciones al diagnóstico médico mediante el aprendizaje profundo aplicado a datos retinianos

Glaucoma is a progressive optic neuropathy and the leading cause of irreversible blindness worldwide. The disease is associated with retinal structural damage as well as functional vision loss. Yet its asymptomatic onset and heterogeneous presentation make diagnosis inherently complex. Current clinical practice relies on a combination of structural imaging, such as Optical Coherence Tomography (OCT) and Color Fundus Photography (CFP), and functional assessment via Standard Automated Perimetry (SAP). These methods are resource-intensive, subjective, and poorly scalable, limiting their effectiveness in routine clinical use. This thesis proposes Artificial Intelligence (AI)-based frameworks that leverage ocular imaging to address these challenges by enabling automated, objective, and scalable approaches for glaucoma diagnosis and functional assessment. To advance automated glaucoma diagnosis from OCT volumes, this thesis introduces a deep learning-based framework for holistic analysis of the entire ocular structure. The proposed architecture integrates a pre-trained Vision Transformer (ViT) for feature extraction from individual B-scans with a Bidirectional Gated Recurrent Unit to capture inter-slice spatial dependencies. This design enables the comprehensive modeling of local retinal details and global structural integrity, both critical for accurate glaucoma diagnosis. Evaluation on a real-world 3D OCT dataset demonstrates superior performance and robustness compared to state-of-the-art slice-based and volumetric methods, underscoring the framework's potential as a clinical decision support tool in glaucoma management. The limitations of SAP for functional assessment have motivated growing interest in predicting Visual Field (VF) indices directly from structural imaging. To this end, this thesis extends the glaucoma classification framework into a multimodal deep learning model that maps complementary information from paired 3D OCT and CFP images to global VF indices, specifically Mean Deviation (MD) and Pattern Standard Deviation (PSD). The architecture comprises two modality-specific encoders for feature extraction and an attention-based fusion module that models cross-modal correlations. Evaluation on two structurally distinct datasets with optic disc and macular fixation points demonstrates improved predictive performance and generalization compared to state-of-the-art approaches. These findings underscore the potential of multimodal structural imaging to enhance functional assessment in glaucoma. In summary, these contributions advance the application of AI in glaucoma by addressing both automated disease diagnosis and functional assessment. Using volumetric OCT analysis and multimodal structure-function mapping, this thesis demonstrates that deep learning can capture subtle and spatially distributed patterns across ocular modalities that are often overlooked in conventional approaches. The proposed frameworks highlight the potential of AI to improve diagnostic accuracy, scalability, and accessibility, supporting clinical decision-making and more equitable glaucoma care.