SgCG: Semantic-guided Contrastive Generalization for Medical Image Segmentation

Mengzhu Wang, houcheng su, Zeheng Wang, Dianxi Shi, Meilin Zhang, Yin Nan, Jiao Li, Bingli Wang, Li Shen

Keywords: Medical Image Segmentation;Semantic-Guided;Contrastive Generalization; Domain-invariant Feature

Abstract:

After training on the source domain, deep learning models often struggle to generalize effectively to unknown target domains with differing data distributions. This is an even more severe challenge when the target domain is not available. In this paper, we tackle the problem of domain-generalized medical image segmentation by introducing a novel semantic-guided contrastive generalization algorithm, termed SgCG. The method aligns different multi-source domains based on semantic distributions to learn domain-invariant features.
Specifically, we implement a novel contrastive generalization loss at the pixel level that incorporates semantic distributions from the source domains. This approach facilitates the clustering of pixel representations from the same category while effectively separating those from different categories, thereby improving the model's segmentation performance while learning domain-invariant features. Furthermore, we establish an upper bound estimation for the SgCG approach by integrating a contrastive generalization loss which include an infinite number of both similar and dissimilar pixel pairs. Despite the simplicity and straightforwardness of the approach, our empirical analysis reveals mechanisms that can maximize the potential of SgCG. We demonstrate the effectiveness of our approach using two public benchmarks for generalizable segmentation in medical images, where it achieves state-of-the-art performance.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1789