Leveraging Deep Learning for Automated Diabetic Retinopathy Detection and Classification with Convolutional Neural Networks

A deep learning-based approach is presented in this study for multi-class classification of DR from a dataset of 2,750 retinal fundus images obtained from Kaggle. The dataset spans five classes: Preprocessed with resizing, normalization, and data augmentation: Healthy, Mild DR, Moderate DR, Severe DR, and Proliferative DR. At the first stage, a combined model is proposed, which employs CNN as the backbone for feature extraction and further fine-tuned with the Dense layer, Batch Normalization, Dropout and a softmax output layer for the multi-class classification. An overall accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$92 \%$</tex> and an F1 score of a weighted average of 0.93 was achieved by training the model using the AdamW optimizer over 50 epochs. Evaluation metrics such as precision, recall, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{F1}$</tex>-score and confusion matrix analysis demonstrate high performance for Healthy and Moderate DR classes but challenging performance on rare classes like Proliferative DR and Severe DR; both F1 scores are 0.84. In particular, limitations arise due to class imbalance and overlapping features. In future work, forward work will involve mitigating these issues by means of advanced feature extraction, data rebalancing, and expanding datasets to increase generalizability. Due to the model's demonstrated high accuracy and computational efficiency, it has shown promise as a fundamental world clinical tool for aiding ophthalmologists in the early, timely diagnosis and management of DR.