Early Detection and Severity Classification of Diabetic Retinopathy Through Image Processing and Deep Learning

Diabetic retinopathy (DR) remains a leading cause of blindness among individuals with diabetes, making early detection crucial to mitigate vision deterioration and prevent vision loss. However, the current manual DR diagnosis by ophthalmologists in clinics is labor-intensive, particularly due to challenges in identifying subtle retinal lesions. Hence, there is a need to develop an automated detection framework. This study leveraged the two pre-trained ResNet50 and VGG19 models for detecting DR in retinal fundus images from the APTOS 2019 dataset. The advanced image preprocessing techniques such as the Graham method, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) were explored. The results demonstrate that combining these preprocessing methods with optimized parameters significantly enhances performance. For binary classification, the pretrained VGG19 model, combined with the proposed CLAHE + ESRGAN preprocessing method, achieved an accuracy of 98.91%. For multi-class classification, the pre-trained ResNet50 model, combined with the proposed Graham20 + CLAHE and ESRGAN + CLAHE preprocessing methods, attained accuracies of 86.07% and 84.15%, respectively. This study is the first systematic study of deep learning architectures with various combinations of DR image preprocessing methods, offering a significant step forward in automated DR detection and severity classification.