Reproduce “Source-free unsupervised domain adaptation for cross-modality abdominal multi-organ segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jin Hong, Yu-Dong Zhang, Weitian Chen
A two-stage source-free unsupervised domain adaptation framework is presented for cross-modality abdominal multi-organ segmentation, enabling knowledge transfer from labeled CT to unlabeled MR images and vice versa without source data access. The method achieved a mean Dice Similarity Coefficient of 0.888 for abdominal organ segmentation in MR, outperforming direct application and several state-of-the-art domain adaptation techniques.
Key insights
- Feature map statistics (FMS)-guided model adaptation is essential for enabling initial source-free knowledge transfer, preventing model collapse by aligning target feature distributions.
- Element-wise style compensation effectively bridges the domain gap at the image level by transforming target images into a source-like style while preserving crucial organ contrast.
- Self-learning mechanisms like Pixel-Adaptive Mask Refinement (PAMR) and a circular learning strategy provide crucial refinements, improving local segmentation consistency and overall accuracy in later training phases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.