Automated Diabetic Retinopathy Severity Classification Using Transfer Learning with DenseNet201

Abstract

Diabetic Retinopathy (DR) remains one of the most common causes of preventable blindness worldwide, particularly among working-age adults between 20 and 64 years of age. Despite its prevalence, early diagnosis continues to be a significant clinical challenge because the disease frequently progresses without noticeable symptoms until substantial and often irreversible damage has already occurred to the retinal tissue. Methods: An automated classification system was developed to identify the severity of Diabetic Retinopathy from fundus images using a convolutional neural network built on the DenseNet201 architecture with transfer learning from ImageNet pretrained weights. The system classifies retinal images into five distinct stages of severity: No DR, Mild Non-Proliferative DR, Moderate Non-Proliferative DR, Severe Non-Proliferative DR, and Proliferative DR. To address the substantial class imbalance present in the training dataset, image augmentation techniques including random horizontal flipping and random rotation were employed. The model was trained on Gaussian-filtered retinal images sourced from a publicly available Kaggle dataset containing 3,662 original samples, with a 75-25 stratified train-test split to preserve class proportions. The model was evaluated using multiple metrics including accuracy, weighted precision, weighted F1 score, and a detailed confusion matrix analysis across all five classes. Results: The experimental results demonstrate that the proposed approach achieves 94% overall accuracy, a weighted precision of 0.949, and a weighted F1 score of 0.937 on the validation set containing 500 randomly sampled images. A thorough analysis of per-class performance is also presented, along with a comparison of results against other popular architectures reported in the literature. The limitations of the current approach are discussed, and concrete directions for future improvement are proposed, including multi-model ensembling, attention mechanisms, and explainability through gradient-based visualization techniques. The complete implementation including data preprocessing, model training, and validation scripts is provided in a public GitHub repository to support full reproducibility of the reported results.