Binary and Multi-Label Ocular Disease Detection from Retinal Fundus Images Using Mobilenetv2 and Resnet50 with Grad-CAM Interpretability and Clinical Web Deployment

Abstract

Ocular diseases such as diabetic retinopathy, glaucoma, cataract, age-related macular degeneration (AMD), and pathological myopia are principal causes of preventable blindness worldwide. Early automated diagnosis is imperative yet remains inaccessible in resource-limited settings. This paper presents a comprehensive deep learning framework for multi-disease ocular screening on the publicly available ODIR-5K dataset, comprising 14,399 annotated fundus images across eight diagnostic categories. Two complementary strategies are implemented: (i) binary, one-versus-rest classification for seven individual diseases using MobileNetV2 with a frozen ImageNet backbone and taskspecific classification heads, and (ii) multi-label classification using a fine-tuned ResNet50 architecture for simultaneous prediction of all eight classes. An extensive preprocessing and augmentation pipeline is applied to enhance generalization under class imbalance. Model performance is assessed via accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). Binary MobileNetV2 classifiers demonstrate superior disease-specific sensitivity and precision relative to the multilabel model. ResNet50 achieves AUC scores of 0.97 for cataract and pathological myopia but struggles with minority classes such as hypertension and AMD. Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps and binary attention masks provide clinically interpretable visual evidence of model decisions. The trained models are integrated into a Streamlit web application offering real-time image upload, ranked probability outputs, class-specific Grad-CAM visualizations, and personalized clinical recommendations, establishing a practical pathway from research to ophthalmic screening.