Diabetic retinopathy (DR) and glaucoma affect more than 400 million people worldwide and are among the leading causes of preventable blindness. Early diagnosis of these diseases is critical to preventing vision loss. However, in regions with limited access to healthcare, the shortage of qualified ophthalmologists makes timely diagnosis of these diseases difficult. In this study, a hybrid deep learning architecture combining the powerful feature extraction capabilities of EfficientNet-B3 and ResNet50 architectures was developed for the automated and accurate diagnosis of diabetic retinopathy and glaucoma from fundus images. The proposed HybridNet architecture adopts a multi-scale feature fusion strategy, simultaneously analyzing both micro-scale lesions (microaneurysms, hard exudate, retinal hemorrhages) and macro-scale structural tissue changes (optic disc geometry, retinal vascular network morphology) in fundus images. Additionally, integrated Grad-CAM heat maps provide clinical interpretability, demonstrating the model's focus on medically critical areas such as pathological hemorrhages, the optic disc region, and vascular structures during the decision-making phase, offering physicians reliable guidance in the diagnostic process. Experimental findings show that the developed system achieved high-performance metrics such as a 98.0% F1-score and 0.99 AUC in the DR class, a 0.972 F1-score in the Glaucoma class, and a 0.978 F1-score in the Healthy class. These results demonstrate that the proposed hybrid architecture can create a reliable screening infrastructure even in regions with limited access to healthcare and can offer transparent and high-performance diagnostic capabilities in clinical decision support systems. In conclusion, this study makes a significant contribution to the field of AI-powered medical image analysis, highlighting the vital role that deep learning technology plays in the diagnosis of retinal diseases and the democratization of healthcare.
Yanık et al. (Tue,) studied this question.