Diabetic Retinopathy Grading Base on Contrastive Learning and Semi-supervised Learning

The diabetic retinopathy (DR) detection based on deep learning is a powerful tool for early screening of DR. Although several automatic DR grading algorithms have been proposed, their performance is still limited by the characteristics of DR lesions and grading criteria, and coarse-grained image-level label. In this paper, we propose a novel approach based on contrastive learning and semi-supervised learning to break through these limitations. We first employ contrastive learning to solve the problem of inter-class and intra-class differences in DR grading. This method enables the model to identify the unique lesion features on each DR fundus color image and strengthen the feature expression for different kinds of lesions. Then we use a small amount of open-source pixel-level annotation dataset to train the lesion segmentation model, in order to provide fine-grained pseudo-label for image-level fundus images. Meanwhile, we design a pseudo-label attention structure and deep supervision method, to increase the attention of the model to lesion features and improve the grading performance. Experiments on the open-source DR grading datasets EyePACS, Messidior, IDRiD, and FGADR can prove the effectiveness of our proposed method and show the results superior to the previous methods.