Beyond Classification: An Explainable Deep Learning Framework for the Early Detection and Severity Grading of Diabetic Retinopathy in Fundus Images

Diabetic Retinopathy (DR) remains a primary cause of preventable vision loss globally, necessitating efficient and accurate screening methodologies. While deep learning models, particularly Convolutional Neural Networks (CNNs), have demonstrated human-level performance in classifying retinal fundus images, their “black-box” nature hinders clinical adoption and diagnostic trust. This paper proposes a robust, explainable deep learning framework designed to detect and grade the severity of DR from fundus photography. We utilize an ensemble of pre-trained architectures fine-tuned on a high-quality clinical dataset, incorporating a feature-attention mechanism to focus on pathological markers such as microaneurysms and hemorrhages. To address the challenge of interpretability, we integrate Gradient-weighted Class Activation Mapping (Grad-CAM) to generate heatmaps that visualize the specific retinal regions influencing the model’s classification. Experimental results demonstrate that our proposed model achieves a high sensitivity and quadratic weighted kappa score, outperforming existing baseline models. Furthermore, qualitative analysis reveals that the attention maps correlate strongly with expert ophthalmologist annotations. Our findings suggest that this framework not only improves automated diagnostic accuracy but also enhances clinical decision support by providing transparent, localized evidence for each diagnosis.

کلیدواژه ها:

: Diabetic retinopathy ، Deep learning ، Convolutional neural networks ، Explainable AI ، Fundus imaging ، Medical image analysis ، Grad-CAM ، Ophthalmology