Domain adaptation for retinal vessel segmentation using asymmetrical maximum classifier discrepancy

Jingyu Zhuang, Ziliang Chen, Junyi Zhang, Dongyu Zhang, Zhaoquan Cai

Abstract: Retinal vessel segmentation is indispensable for ophthalmic computer-aided diagnosis (CAD) and large-scale automatic detection system of retinal diseases. Recently, various convolutional neural networks (CNN) have been applied in retinal vessel segmentation and obtain impressive performance. However, due to the domain shift caused by the variation between source and target datasets, these methods poorly generalize across different datasets. In this paper, we develop asymmetrical maximum classifier discrepancy (AMCD) approach from maximum classifier discrepancy for unsupervised domain adaptation. We utilize labeled source data and unlabeled target data to train a model and test it on the target domain. In our work, we use three classifiers asymmetrically which means that two assist classifiers are used to maximize the discrepancy on target samples and one main classifier is trained only by the source samples. We have validated our approach on DRIVE, STARE, CHASE-DB1 and IOSTAR eye vessel segmentation datasets and the experimental results showed that our method can achieve significantly accuracy in the setting of unsupervised domain adaptation.