Intelligent Diagnosis of Chronic Glaucoma and Its Differentiation from Diabetic Retinopathy Using Machine Learning

Abstract

Early detection of DR and Chronic Glaucoma is crucial to prevent permanent vision loss, as they don't have symptoms at their initial stages. However, due to the overlapping characteristics of the retina and the need for clinical skill to interpret these, it remains difficult to differentiate between the various diseases of the eye. This study proposes an intelligent multi-task deep learning framework to automatically diagnose, grade and identify DR and chronic glaucoma based on retinal fundus images and structured clinical criteria. The proposed architecture consists of an attention mechanism to improve clinically relevant regions like the optic disc, optic cup, and the retina blood vessels, and a transfer learning-based EfficientNet as a backbone to learn discriminative features in the retina. A multimodal data fusion approach is used, where deep image features are fused with clinical parameters, e.g., blood glucose levels and intraocular pressure, to boost diagnostic accuracy. The feature representation is integrated and used for simultaneously solving the problem of illness classification, severity evaluation and anatomical localization in a single multi-task learning architecture. The classification accuracy, precision, recall, and F1-score of the proposed approach are better than the conventional CNN models with retinal fundus image datasets made public. Based on experimental results, the proposed paradigm enhances the interpretability and robustness of the classification, and it demonstrates a good ability to differentiate diabetic retinopathy from chronic glaucoma, in a diagnostic way. The resulting system is a reliable, scalable and efficient computer-aided diagnostic system that can assist ophthalmologists in large scale screening programs, diagnose disease early, and facilitate clinical decisions regarding retinal diseases that are a threat to vision.