Answers are grounded in this paper and cite their sources.
Ocular diseases, including Diabetic Retinopathy (DR), Glaucoma, Pterygium, and Retinal Detachment, require early detection to prevent irreversible blindness. However, manual screening is labor-intensive, and standard deep learning models often fail to capture the complex global spatial relationships necessary for multipathology grading. To address these limitations, we propose a MobileNetV2-Transformer Hybrid (MoViT-Eye) architecture for the automated 5-class classification of fundus images. This framework integrates the efficient local spatial feature extraction of MobileNetV2 with the long-range dependency modeling of Transformers. The study uses 4,906 cleaned images and applies LIME, Grad-CAM++, and Saliency Maps to identify image regions that influence each prediction. These methods generate heatmaps that highlight important clinical features, including microaneurysms and optic disc abnormalities, and support transparent clinical decision-making. The proposed model significantly outperforms existing architectures, achieving an accuracy of 90.36%, a Matthews Correlation Coefficient (MCC) of 0.8643, a Cohen’s Kappa of 0.8640, and a specificity of 97.29%. The model achieved an Area Under the Curve (AUC) of 98.60% for identifying DR and Retinal Detachment. By combining hybrid feature learning with multi-layered interpretability, this model provides a robust and transparent tool for automated clinical ophthalmic screening.
No takes yet. Share an insight, caveat, or question.
Hridoy et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: