Abstract: Early identification of ocular diseases is essential to preventing irreversible vision loss, particularly among diabetic patients who are highly susceptible to multiple retinal complications. Although several automated approaches have been developed in recent years, most are limited to single-disease detection or depend heavily on conventional segmentation and fixed-parameter models, which often fail to capture subtle and overlapping abnormalities in fundus images. To address these gaps, this research proposes an integrated framework that combines adaptive segmentation, intelligent optimisation, and attention-driven deep learning for multi-class ocular disease classification. The study begins by gathering retinal fundus images from widely used public repositories and enhancing them to improve structural visibility. An Adaptive SegUNet model is then employed to segment clinically relevant regions, while its parameters are automatically refined using the Improved Mother Optimization Algorithm to achieve more stable and precise segmentation. The segmented images are subsequently analysed through an Attention-based Efficient Deep Learning network incorporating a Gated Recurrent Unit enabling the model to learn discriminative spatial and sequential patterns associated with different ocular disorders. Finally, the performance of the proposed system is evaluated and compared with existing methods to demonstrate its effectiveness. The results highlight the potential of the ASUNet–IMOA–AEDGRU pipeline as a reliable and robust solution for multi-disease retinal screening, offering valuable support for clinical diagnosis and early intervention.

Keywords: Adaptive SegUNet (ASUNet), Improved Mother Optimization Algorithm (IMOA), Attentionbased Deep Learning, Gated Recurrent Unit (GRU), Retinal Fundus Images, Ocular Disease Classification, Medical Image Segmentation, Multi-class Diabetic Eye Disease Detection, Computer-Aided Diagnosis

Pages: 89-114

WSEAS Transactions on