An Optimized Deep Learning Framework with Transfer Learning for Diabetic Retinopathy Detection

Diabetic Retinopathy (DR) is a leading disease worldwide. It causes vision impairment and can lead to permanent vision loss. Precise and early detection of DR grading is essential to prevent disease progression and blindness. The traditional screening process is manual diagnosis from retinal fundus images by clinicians. However, this manual process takes a lot of time, and the diagnosis may be subjective. In the medical field, accurate diagnosis is important for future treatment and patient care. When permanent damage to human organs is possible, diagnosis should be especially accurate. An automated system can provide unbiased and correct results for classifying DR severity. This study focuses on developing deep learning models to classify DR severity by using open-source retinal fundus images. The deep learning models—Convolution Neural Network (CNN), Inception V3, and ResNet-50—were implemented with optimizations. Among these three models, ResNet-50 achieved the highest accuracy of 90.3%. CNN and Inception V3 models achieved accuracies of 85.8% and 88.6%, respectively. This study also compares other classification metrics such as precision, recall, and F1-score. Cohen’s kappa was also evaluated for all three models. ResNet-50 achieved a Cohen’s kappa of 0.87. This result shows that the proposed model's performance in DR classification is acceptable.