Comparison of ResNet-50 and DenseNet-121 Architectures in Classifying Diabetic Retinopathy

Diabetic Retinopathy (DR) is a severe complication of diabetes that can lead to blindness if not detected early. However, early diagnosis remains difficult due to subtle initial symptoms. With advancements in technology, deep learning has become a reliable approach for DR classification through retinal image analysis. This study compares the performance of two CNN architectures, ResNet-50 and DenseNet-121, in classifying DR using 2,000 retinal images categorized into four classes: normal, mild, moderate, and severe. Models were trained using both Stratified K-Fold Cross Validation (k=5) and non-K-Fold approaches. Data augmentation techniques, including rotation, flipping, zooming, and translation, were applied to enhance model generalization. Training was optimized using the Adam optimizer with a learning rate of 0.001, ReduceLROnPlateau for adaptive learning rate adjustment, and a dropout rate of 0.2 to prevent overfitting. Results showed that ResNet-50 outperformed DenseNet-121, achieving 84% accuracy (without K-Fold) and 83% (with K-Fold), while DenseNet-121 reached 80% and 81%, respectively. ResNet-50 also yielded higher precision, recall, and F1-score across all classes. Overall, the combination of a robust architecture, proper regularization, and optimization significantly improved classification performance and reduced overfitting.