Classification of retinal fundus images with deep learning for diagnosis of diabetic retinopathy

In this study, the diagnosis of diabetic retinopathy and the determination of the level of the disease is considered as a classification problem and prediction is attempted. For this purpose, the 224x224 filtered version of the APTOS 2019 dataset, which is a popular dataset, was selected and a total of 5 new datasets were created from this dataset, including multi-class and low-class datasets. A large-scale modeling study was conducted by applying CNN-based architectures (5 Layer CNN, EfficientNetB0, EfficientNetB1, VGG16, VGG16-TL, VGG19-TL, MobilNet, MobileNet V2, our designed CNN, GoogleNet, DenseNet121 and ResNet50) to the original dataset and the newly created new datasets and by choosing different data partitions (80-20 training-testing, 3,5 and 10 times cross-validation). As a result, the best model for disease level classification was found to be the MobilNet V2 model with 94.82% test accuracy and the best model for disease prediction was found to be the DenseNet121 model with 99.26% test accuracy.