Optimizing Deep Learning Architectures for Diabetic Retinopathy Detection: A Comparative Study of CNN and ResNet-50 With ADAM and SGDM

Diabetic Retinopathy (DR) is a serious complication of diabetes and a major cause of vision loss. The damage it causes to the retina is permanent, which makes early detection and accurate grading essential to prevent blindness. This study evaluates the performance of two deep learning models using two different optimizers for automatic screening and severity grading of DR from retinal fundus images. The models compared are a Convolutional Neural Network (CNN) and a Residual Network (ResNet-50), trained with Adaptive Moment Estimation (ADAM) and Stochastic Gradient Descent with Momentum (SGDM). Two public datasets, APTOS and EyePACS, were used. Images were preprocessed using Contrast Limited Adaptive Histogram Equalization (CLAHE) for contrast enhancement and data augmentation for class balance. Performance was evaluated using accuracy, area under the curve (AUC), F1-score, recall, and precision. ResNet-50 with the ADAM optimizer achieved the best results, with an accuracy of 80.71% and an AUC of 0.9539 on the APTOS dataset. These results show that ResNet-50 with ADAM is an effective model for automated DR screening and can support early detection and prevention of diabetes-related vision loss.