Diabetic Retinopathy Classification Using Deep Learning

Diabetic Retinopathy (DR) is still one of the main causes of vision impairment, thus early-stage diagnosis is important to ensure timely intervention. This work discusses deep learning-based DR severity classification using convolutional neural networks (CNNs), such as MobileNet, EfficientNetB0, NASNetMobile, and VGG16. Diabetic retinopathy fundus images were preprocessed through normalization and augmentation strategies to improve the performance of models. All architectures were pre-trained and hyperparameter-tuned for optimal classification accuracy. Out of the models that were tested, EfficientNetB0 scored the highest accuracy at 87%, followed by MobileNet with an accuracy of $\mathbf{8 1 \%}$ and NASNetMobile at $\mathbf{7 4 \%}$. VGG16 scored an accuracy of 82%. All the models worked well in identifying proliferative DR but identifying between the mild and moderate stages was difficult because retinal features overlapped. These results show the suitability of CNN-based systems for real and reproducible DR classification. The developed models are capable of aiding clinical decision-making and facilitating early treatment, particularly in limited-resource environments.