Reproduce “Unsupervised Domain Adaptation via Similarity-based Prototypes for Cross-Modality Segmentation” on OpenResearch

About this paper

Ziyu Ye, Chen Ju, Chaofan Ma, Xiaoyun Zhang

Researchers from Shanghai Jiao Tong University developed an unsupervised domain adaptation framework for cross-modality medical image segmentation, leveraging similarity-based prototypes and contrastive learning. The method achieved an average Dice score of 77.5% on the MRI to CT cardiac segmentation task, outperforming previous state-of-the-art UDA approaches.

Key insights

• Explicitly enforcing semantic consistency through similarity-based prototypes and contrastive learning provides more robust cross-domain feature alignment than implicit adversarial methods.
• Maintaining feature dictionaries to collect prototypes from multiple samples effectively mitigates the "class-missing problem" common in single-image prototype generation, enhancing the reliability of prototype-based learning.
• Regularizing both intra-class feature consistency and inter-class prototype separability simultaneously improves the discriminative power of the embedded feature space, leading to better segmentation across modalities.

Summary by alphaXiv.