Towards Semi-Supervised Segmentation of Retinal Fundus Images via Self-Training

Glaucoma is an eye disease that may cause blindness by damaging the optic nerve. The optic cup-to-disc ratio is one of the most important criteria in the diagnosis of glaucoma. However, accurately partitioning a retinal fundus image into optic cup and optic disc regions is crucial to precisely estimating the cup-to-disc ratio automatically. With the emergence of Deep Neural Networks (DNN) and available large-scale manually labeled training data, generic image segmentation has made great progress in recent years. However, large-scale well-labeled medical images are usually expensive and difficult to obtain. To address this problem, in this paper we propose a semi-supervised learning method for retinal fundus image segmentation via self-training based on the MR-Net. The proposed approach uses a self-training semi-supervised learning framework to generate pseudo-labels for unlabeled images. To improve the accuracy of the pseudo-labels, a dense conditional random field is introduced to refine the generated pseudo-labels during the self-training process. Experimental results show that the proposed method remarkably achieves state-of-the-art performance on the RIGA benchmark using only 50% of the annotated data for training, well alleviating the shortage of annotated training data in retinal fundus image segmentation.