A Deep Learning Model Comparation for Diabetic Retinopathy Image Classification

Purpose: This study compares the performance of various deep learning models for diabetic retinopathy (DR) classification, emphasizing the impact of different optimization functions. Early detection of DR is vital for preventing blindness, and the research investigates how optimization functions influence the classification accuracy and efficiency of several convolutional neural networks (CNNs). This study fills a gap in the existing literature by examining how optimization functions affect model performance in conjunction with architectural considerations. Methods: This paper uses the APTOS 2019 dataset, which comprises 3,663 retinal fundus images classified into five classes of diabetic retinopathy severity. Four CNN-based models, including CNN, ResNet50, DenseNet121, and EfficientNet B0, were trained using five optimization techniques: Adam, SGD, RMSProp, AdamW, and NAdam. The performance of the experimental scenarios was evaluated through accuracy, precision, recall, F1-score, training duration, and model size. Result: EfficientNet B0 demonstrated superior computational efficiency with a minimal model size of 16.16 MB. Subsequently, DenseNet121 with the SGD optimizer achieved the highest test accuracy of 96.86%. The experimental results indicate that the optimizer significantly influences model performance. AdamW and NAdam yield superior outcomes for deeper architectures such as ResNet50 and DenseNet121. Novelty: This paper offers an analytical examination of deep learning models and optimization techniques for DR classification, helping to clarify the trade-offs between computational efficiency and classification performance. The findings contribute to the development of more accurate and efficient DR detection systems, which could be utilized in real-world, resource-limited settings.