Ocular diseases such as diabetic retinopathy, glaucoma, and cataract remain major causes of vision loss worldwide, creating an increasing demand for accurate and reliable automated diagnostic systems. Although deep learning techniques have demonstrated strong performance in retinal image analysis, their clinical adoption is still limited by challenges related to predictive reliability, robustness, and interpretability. This study presents HEFR-Net (Hybrid EfficientNet-B3 and ResNet-50 Feature-Fusion Network), a hybrid deep learning framework designed for multi-class classification of retinal fundus images. The proposed architecture integrates EfficientNet-B3 and ResNet-50 through feature-level fusion, enabling the model to capture both fine-grained texture patterns and high-level structural characteristics associated with retinal diseases. To enhance predictive reliability, label smoothing is incorporated as a calibration-oriented regularization strategy, reducing overconfidence in model predictions. In addition, Grad-CAM is employed to provide visual explanations by highlighting clinically relevant regions that contribute to the model's decisions. The framework is evaluated on a publicly available dataset consisting of 4,217 retinal fundus images across four diagnostic categories: Normal, Cataract, Glaucoma, and Diabetic Retinopathy. Experimental results show that HEFR-Net achieves an accuracy of 95.50%, indicating improved performance compared to conventional single-model and standard ensemble approaches. Further analysis demonstrates consistent performance across classes, along with interpretable activation patterns aligned with clinically meaningful features. Overall, the proposed framework provides an accurate, reliable, and interpretable approach for automated retinal disease classification, supporting its potential application in computer-aided diagnosis and clinical decision support systems.
Duhaim et al. (Mon,) studied this question.