Hybrid Graph Convolutional Network for Semi-Supervised Retinal Image Classification

Diabetic Retinopathy (DR) causes a significant health threat to the patient's vision with diabetic disease, which may result in blindness in severe situations. Various automatic DR diagnosis models have been proposed along with the development of deep learning, while there always relies on a large scale annotated data to train the network. However, annotating medical fundus images is cost-expensive and requires well-trained professional doctors to identity the DR grades. To overcome this drawback, this paper focuses on utilizing the easily-obtained unlabeled data with the help of limited annotated data to identify DR grades accurately. Hence we proposes a semi-supervised retinal image classification method by a Hybrid Graph Convolutional Network (HGCN). This HGCN network designs a modularity-based graph learning module and integrates Convolutional Neural Network (CNN) features into the graph representation by graph convolutional network. The synthesized hybrid features are optimized by a semi-supervised classification task which is assisted by a similarity-based pseudo label estimator. Through the proposed HGCN method, the retinal image classification model can be trained efficiently by partially labeled samples and the complicated annotating work is not required for the most retinal images. The experimental results on MESSIDOR dataset demonstrate the favorable performance of HGCN on semi-supervised retinal image classification, and the fully labeled data training also achieves an obvious superiority to the state-of-the-art supervised learning methods.