Enhancing cross-domain generalization in retinal image segmentation via style randomization and style normalization

Abstract

Retinal image segmentation is a crucial procedure for automatically diagnosing ophthalmic diseases. However, existing deep learning-based segmentation models suffer from the domain shift issue, i.e., the segmentation accuracy decreases significantly when the test and training images are sampled from different distributions. To overcome this issue, we focus on the challenging single-source domain generalization scenario, where we expect to train a well-generalized segmentation model on unseen test domains with only access to one domain during training. In this paper, we present a style randomization method, which performs random scaling transformation to the LAB components of the training image, to enrich the style diversity. Furthermore, we present a style normalization method to effectively normalize style information while preserving content by channel-wise feature standardization and dynamic feature affine transformation. Our approach is evaluated on four types of retinal image segmentation tasks, including retinal vessel, optic cup, optic disc, and hard exudate. Experimental results demonstrate that our method achieves competitive or superior performance compared to state-of-the-art approaches. Specifically, it outperforms the second-best method by 3.9%, 2.6%, and 4.8% on vessel, optic cup, and hard exudate segmentation tasks, respectively. Our code will be released at https://github.com/guomugong/SRN.