Retinopathy Diagnosis Using Semi-Supervised Multi-Channel Generative Adversarial Network

Various kinds of retinopathy are the leading causes of blindness in human being, and with the rapid development of fundus images (FI) analysis in recent years, deep learning has became the focus while using Computer-Aided-Diagnosis (CAD) system to diagnose retinopathy. However, the conventional deep learning method usually rely on sufficient number of labeled FI, but the cost of acquiring enough labeled FI is too expensive, in addition the difficulty in identifying the tiny lesions and the randomness of lesions distribution also brings great challenge to CAD via deep learning. Therefore, this paper proposes a semi-supervised multi-channel generative adversarial network (GAN), which can reasonably utilize the unlabeled images and generate new samples to reduce the dependence on the labeled images, meanwhile we introduce into a general feature extraction strategy to avoid the problem of image valid information disappearance caused by downsampling, and improve the robustness of the generative and discriminant network. The experimental results show that our proposed network boosts the classification accuracy by 10.5% compared with the control network using conventional method and reaches the highest of 88.9%.