AN EXPLAINABLE DEEP RULE BASED FOR DIABETIC RETINOPATHY CLASSIFICATION

Abstract

Diabetes is a chronic disease that affects millions of people worldwide. Among its most serious complications is Diabetic Retinopathy (DR), an eye condition that can lead to vision loss and even blindness in the absence of early diagnosis. Traditional screening methods rely on the manual examination of fundus images, a procedure that is often time-consuming and prone to human error. These complications have encouraged researchers to develop automated diagnostic systems based on Artificial Intelligence and Deep Learning. In this thesis, we propose an automated Diabetic Retinopathy detection system based on the EfficientNet-B3 architecture and transfer learning. The model was trained on a combined EyePACS-APTOS-Messidor dataset containing 53,411 balanced fundus images. Advanced preprocessing techniques, such as image resizing to 288×288 pixels, normalization using ImageNet statistics, and data augmentation, were applied. The obtained results show excellent performance with a Quadratic Weighted Kappa score of 0.9134 and an overall accuracy of 85.01%, demonstrating the effectiveness of the proposed approach for early clinical diagnosis.