Integrated Deep Learning Framework for Retinal Disease Detection Using Advanced CNN and Imaging Analysis

Early and accurate detection of diabetic retinopathy (DR) is critical to prevent vision loss in diabetic patients. This study proposes Transfer Learning with Pre-trained Convolutional Neural Network Models (TL-Pretr-CNN), a deep learning-based framework for automated DR detection, severity grading and clinical decision support using high-resolution retinal fundus images. The framework leverages transfer learning with pre-trained convolutional neural network (CNN) backbones (ResNet50, EfficientNet) to extract hierarchical retinal features, while fine-tuning the top convolutional layers and a custom classification head to adapt to domain-specific pathological patterns. Preprocessing techniques, including normalization, contrast-limited adaptive histogram equalization (CLAHE) and data augmentation (rotation, flipping, zoom), enhance lesion visibility and address class imbalance. Gradient-weighted class activation mapping (Grad-CAM) heatmaps provide explainable predictions by highlighting clinically relevant regions, supporting teleophthalmology workflows. The model was trained and evaluated on a large-scale dataset of 130,993 retinal images using stratified sampling (70% training, 15% validation, 15% testing) and its performance was averaged over multiple runs to ensure robustness. TL-Pretr-CNN achieves 88.6% accuracy, 0.78 LLP and 0.85 severity grading concordance, significantly outperforming baseline architectures (Improved Deep Retinal Intelligence Detector Network, DeepMedic, EfficientNet + Grad-CAM) with p < 0.05. Average inference time is ~12 ms per image on an NVIDIA A100 GPU, making the framework suitable for real-time clinical deployment. This work demonstrates that TL-Pretr-CNN can provide reliable, interpretable and clinically relevant predictions, facilitating early DR detection, patient triaging and integration into decision support systems for ophthalmologists. Future work includes validation on multi-institutional datasets and deployment in real-world teleophthalmology environments.