An Adversarial Collaborative-Learning Approach for Corneal Scar Segmentation with Ocular Anterior Segment Photography

Abstract

Corneal scarring is a common eye disease that leads to reduced vision. An accurate diagnosis and segmentation of corneal scar is a critical in ensuring proper treatment. Deep neural networks have made great progress in medical image segmentation, but the training requires large amount of annotated data. Pixel-level corneal scar can only be annotated by experienced ophthalmologists, but eye structure annotation can be done easily by people with minimal medical knowledge. In this paper, we propose Dual-Eye-GAN Net (DEGNet), an end-to-end adversarial collaborative-learning corneal scar segmentation model. DEG-Net can improve segmentation quality with additional data that only has eye structure annotation. We collect the first corneal scar segmentation dataset in the form of anterior ocular photography. Experimental results demonstrate superiority to both supervised and semi-supervised approaches. This is the first empirical study on corneal scar segmentation with anterior ocular photography. The code and dataset can be found in https://github.com/kaisadadi/Dual-GAN-Net.