An Automated Diagnosis of Diabetic Retinopathy Grading Using DenseNet169

This study used DenseNet121, DenseNet169, DenseNet201, EfficientNetB0, EfficientNetB7, Inception ResNetV2, Inception V3, MobileNet, MobileNetV2, MobileNetV3Large, ResNet50, ResNet101 and ResNet152, 13 different deep learning neural network models were used for automatic grading of diabetes retinopathy(DR). The change of Validation Accuracy trend was initially observed with 10, 20, 30, 40 and 50, different EPOCHs. Before grading, pre-processing such as elimination, amplification, denoising, data enhancement, normalization, etc. were carried out for the problems in the retinal image data set in the experiment. Then we used the deep learning neural network model proposed in this study to embed the attention mechanism into DenseNet169 for automatic grading of DR, enable deep learning neural network models to focus on features such as exudates, thick blood vessels and microaneurysms in retinal images. In order to avoid vanishing gradients in the neural network model and ensure classification accuracy, DenseNet169 was adjusted by using a targeted swish function for processing, and classify retinal images into three levels and finally the classification accuracy of the neural network model proposed in this study reached 99.76%, macro avg precision reached 99.76%, macro avg recall reached 99.76% and macro avg f1-score reached 99.76%. It is significantly better than other models and on average. It can correctly identify and automatically grading more than 4600 images in the verification of 4,617 retinal images. The method proposed in this study has been validated to effectively improve the accuracy of DR classification and can be used for large-scale DR medical imaging screening. It can effectively assist doctors in achieving efficient diagnosis of DR grading. This work was supported by Research Program of Qilu Institute of Technology (No.QIT24NN032).