Improved Multiclass Classification of Eye Diseases using a Feature-Augmented Enhanced Deep Learning Approach

Alvin Ming Siang, Choo (2026) Improved Multiclass Classification of Eye Diseases using a Feature-Augmented Enhanced Deep Learning Approach. Masters thesis, Universiti Malaysia Sarawak (UNIMAS).

PDF (OPEN ACCESS) Alvin Choo Ming SiangMSc Thesis.pdf Available under License Creative Commons Public Domain Dedication. Download (6MB)

Abstract

Vision impairment is a global health issue often caused by conditions such as cataracts, diabetic retinopathy, and glaucoma. Early detection is critical to prevent irreversible vision loss. Retinal fundus imaging plays a central role in diagnosis, and deep learning offers promising automation for disease detection. However, multiclass eye disease classification remains challenging due to limited annotated datasets, overlapping clinical features and intra-class heterogeneity. This research proposes an enhanced deep learning pipeline for classifying retinal fundus images into four classes: cataracts, diabetic retinopathy, glaucoma, and normal. A dataset, named CDGN, was constructed by integrating and standardizing images from eight publicly available sources, offering improved diversity, resolution, and patient demographics to enhance model robustness and generalizability. Image enhancement was applied to make disease-specific feature more pronounced, while attention mechanisms improved focus on relevant regions, and ensemble learning further boosted performance across heterogeneous data. Multiple convolutional neural network (CNN) architectures were explored through transfer learning. An ablation study quantified the individual contributions of image enhancement, attention, and ensemble learning. Experimental results demonstrate progressive improvements in accuracy, recall, precision, F1-score and AUC across enhancement stages, with the ensemble model achieving the highest performance. These findings indicate that the feature-enhanced deep learning pipeline effectively addresses challenges in multiclass eye disease classification, supporting clinical decision-making and advancing automated diagnostic systems in ophthalmology.

Actions (For repository members only: login required)

View Item