Review of the performance of diabetic retinopathy image classification using convolutional neural network models

This study examines the critical importance of early detection in diabetic retinopathy (DR), a prevalent condition among people with diabetes that can lead to vision loss. It explores how convolutional neural networks (CNNs) can help identify key features in fundus images to detect early signs of DR. CNNs are powerful tools for processing unstructured data, especially images, skillfully decoding intricate data for DR detection. This article presents the architecture of CNNs, including input, convolutional, pooling, and fully connected layers, and examines two prevalent CNN architectures: Inceptionv3 [1][2] and ResNet34 [3][4][5]. These architectures were evaluated using two public diabetic retinopathy datasets, totaling 37,770 images at different stages of various resolutions and quality. Instances of ResNet34 and InceptionV3 models were created, each trained with both datasets. The results reveal that the accuracy of the ResNet34 model trained with dataset 1 ranges from 63.79% to 95.93%, while the ResNet34 model trained with dataset 2 performs between 73.04% and 75.37%. Regarding InceptionV3, its accuracy on dataset 1 varies from 53.77% to 74.44%, and on dataset 2, accuracy ranges from 71.76% to 77.61%. The results suggest that the ResNet34 model trained with dataset 2 is more robust and generalizable compared to other models, indicating its effectiveness in detecting diabetic retinopathy where image conditions vary.