Deep Learning-Based Diabetic Retinopathy Grading: A Unified Benchmark of CNN, GNN, And RNN Hybrid Architectures

Abstract

Diabetic retinopathy (DR) is a major diabetes-related microvascular complication and one of the leading causes of preventable vision loss. This revised paper presents a notebook-consistent comparative evaluation of six architectures for five-class retinal fundus image grading: DenseNet121 with a support vector machine classifier (DenseNet121+SVM), end-to-end DenseNet121, AlexNet, MobileNetV2 with a graph neural network (MobileNet+GNN), MobileNetV2 with a recurrent neural network (MobileNet+RNN), and a TensorFlow/Keras implementation of standalone MobileNet. The revised experimental description is aligned with the uploaded notebooks, which use local class-wise directories with 506 training images and 217 validation images across Mild, Moderate, No_DR, Proliferate_DR, and Severe classes. The strongest validated result is obtained by end-to-end DenseNet121, which reaches approximately 88.5% validation accuracy, followed by MobileNet+GNN (83.4%), MobileNet+RNN (78.8%), AlexNet (77.0%), standalone MobileNet evaluation accuracy (69.1%), and DenseNet121+SVM (62.7%). The revised Results section is reorganized model-wise and includes confusion matrices, confusion-matrix interpretations, Grad-CAM visual explanations where available, expected/predicted sample outputs, and an explicit discussion of unsupported or inconsistent outputs. The analysis shows that end-to-end fine-tuning is more effective than frozen feature extraction, while lightweight MobileNet hybrids improve over the standalone MobileNet baseline. However, the notebooks also reveal reproducibility issues that must be corrected before final publication, including an unused sixth output node in some PyTorch models, provisional MobileNet confusion-matrix alignment, and absent SVM-specific heat-map evidence.