Intelligent Retinal Image Analysis for Medical Biometrics applications

Abstract

The rapid integration of artificial intelligence into medical imaging has created new op portunities for improving both clinical diagnosis and secure patient identification. Among medical imaging modalities, retinal fundus photography is particularly valuable due to the uniqueness, stability, and diagnostic richness of the retinal vascular network. Retinal im ages simultaneously encode highly discriminative biometric patterns and clinically significant biomarkers associated with ocular and systemic diseases. However, existing studies typically address biometric recognition and pathology classification as separate tasks, limiting the development of unified intelligent systems. This thesis proposes an integrated framework for intelligent retinal image processing in med ical biometric applications, combining deep learning, evolutionary hyperparameter optimiza tion, and multitask learning. A robust retinal authentication pipeline is developed, including advanced preprocessing, residual attention U-Net vessel segmentation, Gaussian-weighted patch reconstruction, and multi-modal feature extraction incorporating morphological, topo logical, and texture-based descriptors. Furthermore, hyperparameter optimization strate gies are systematically evaluated, demonstrating the effectiveness of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) in achieving improved convergence and general ization compared to conventional methods. Finally, a unified multitask architecture with dual backbones and attention mechanisms is introduced to simultaneously perform biometric identification and multi-label pathology classification. Experimental validation on multiple public and clinical datasets confirms high recognition accuracy, strong robustness to image perturbations, and competitive diagnostic performance. The results demonstrate that retinal fundus images can serve as dual-purpose data sources, supporting secure authentication and automated clinical assessment within a single intelli gent framework