Diabetic retinopathy (DR) is among the diabetes related eye conditions that account for much of the world’s preventable sight loss. Whether treatment arrives in time, and which treatment is chosen, depends on how early the condition is caught and how reliably its stage is graded. This work presents a clinical decision support system, built on explainable artificial intelligence (XAI), that grades DR severity directly from colour fundus photographs. Two pre-trained backbones, ResNet50 and EfficientNetB0, are fine-tuned to sort images into the five clinical grades, from an unaffected retina through the mild, moderate and severe non-proliferative stages to proliferative disease. Gradient-weighted class activation mapping is layered on the classifier, so every prediction carries a map of the regions that drove it. For additional risk assessment the framework also considers patient-level variables such as age, blood pressure, blood glucose and glycated haemoglobin. A web based application supports image submission, patient information management, model inference and generation of diagnostic reports. On the APTOS 2019 dataset the EfficientNetB0 grader attains 78.7 percent accuracy and a weighted F1-score of 0.777, improving on the ResNet50 reference, while external testing on MESSIDOR-2 confirms that the behaviour carries across acquisition conditions. The results indicate that the framework can support automated diabetic retinopathy screening while providing interpretable predictions and additional clinical risk information.

Diabetic Retinopathy Retinal images,deep learning,resnet 50,EfficientNetB0,Grad-CAM,Explainable Artificial Intelligence (XAI),clinical decision support system