Multi-class diabetic retinopathy severity classification using Inception-V3 transfer learning with SGDM and Nesterov Accelerated Gradient

Diabetic retinopathy (DR) is a major cause of vision impairment worldwide, early and accurate detection is critical for effective clinical intervention. This study proposes a practical transfer-learning framework to automatically classify severity of diabetic retinopathy in multiple classes using retinal fundus images from the APTOS 2019 dataset. This work aims at investigating a strong and computationally efficient pipeline based on Inception-V3 CNN. In the proposed framework, the data augmentation, transfer learning, stochastic gradient descent with momentum (SGDM), and Nesterov acceleration are combined with image pre-processing to obtain stable training and good classification results. Classification accuracy and macro-area under the receiver operating characteristic curve (macro-AUC) are shown to be 93.45% and 0.9917, respectively, with experimental results over five diabetic retinopathy severity classes. The comprehensive performance evaluation includes a detailed analysis of the confusion matrix, ROC curves, and Grad-CAM visualizations in a class-wise evaluation to further improve the interpretability of the models, thereby aiding clinical understanding of the prediction. The results show that with transfer learning using Inception-V3, Diabetic retinopathy grading with reasonable computational efficiency for moderate-scale screening applications is possible and competitive. But external testing on independent datasets and imaging environments needs to be done before it can be considered for widespread clinical use.