DAPR-Net: Domain Adaptive Predicting-Refinement Network for Retinal Vessel Segmentation

Abstract

The convolutional neural networks (CNN) have been used in various medical image segmentation tasks. However, the training of CNN extremely relies on large amounts of sample images and precisely annotated labels, which is difficult to get in medical field. Domain adaptation can utilize limited labeled images of source domain to improve the performance of target domain. In this paper, we propose a novel domain adaptive predicting-refinement network called DAPR-Net to perform domain adaptive segmentation task on retinal vessel images. In order to mitigate the gap between two domains, the Contrast Limited Adaptive Histogram Equalization (CLAHE) is employed in the preprocessing operations. Since the segmentation result generated by only predicting module can be affected by domain shift, refinement module is used to produce more precise segmentation results and further reduce the harmful impact of domain shift. Atrous convolution is also adopted in both predicting module and refinement module to capture wider and deeper semantic features. Our method has advantages over previous works based on adversarial networks, because in our method smoothing domain shift with preprocessing has little overhead and the data from target domain is not needed when training. Experiments on different retinal vessel datasets demonstrate that the proposed method improves accuracy of segmentation results in dealing with domain shift.