A Diabetic Retinopathy Classification method based on improved InceptionV3 neural network model

With machine learning and deep learning showing good results in feature extraction, image recognition, image classification and other aspects, relevant methods are also increasingly widely applied in medical intelligent diagnosis. Fundus images of diabetic retinopathy (DR) have clear grading standards and obvious lesion characteristics. The study of DR Automatic grading detection is significant for the study of other eye diseases, such as glaucoma and macular degeneration. Therefore, a training classification model DR-ClassifyNet based on the improved InceptionV3 module was proposed, which divided 2D fundus images into five levels to achieve the purpose of assisting diagnosis. In this paper, the traditional digital image processing and deep learning methods are combined. The image was preprocessed by the method of image processing, the vascular structure was segmented by Canny edge detection, and the structure was weakened and removed to facilitate accurate segmentation of focal features. The image was enhanced by adaptive histogram equalization and image mixing with limited contrast. For the training classification model, the improved InceptionV3 module as the core is combined with the attention-mechanism module CBAM, the Dropout method is used to enhance the data to prevent overfitting, the LeaKyReLu activation function increases the nonlinear relationship of the model, and the Softmax function completes the classification. The experimental results showed that the accuracy of the network model for the classification of diabetic retinopathy was as high as 91%, which was significantly higher than that of LeNet, AlexNet and ResNet50 models, and could be used as a reference for other researchers.