Application of Deep Learning Neural Network Architectures in Ocular Disease Classification Models

Abstract

Early diagnosis of ocular diseases is critical to prevent vision impairment, yet traditional methods rely heavily on specialists' expertise and face resource allocation challenges. This study systematically evaluates the performance of diverse deep learning architectures (including CNNs, Transformers, and lightweight models) for multi-class ocular disease classification. Utilizing public ocular image datasets with transfer learning and data augmentation to address class imbalance, our experiments demonstrate that self-attention-based models achieve superior accuracy for complex conditions (e.g., diabetic retinopathy), while lightweight architectures significantly improve computational efficiency. The findings provide empirical guidelines for clinical decision-support systems and propose optimization strategies for architecture design, contributing to practical AI-driven healthcare solutions.