BCMDA: Bidirectional correlation maps domain adaptation for mixed domain semi-supervised medical image segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Mixed domain semi-supervised medical image segmentation (MiDSS) faces challenges due to domain shift and limited annotations.
• Distributional differences impede knowledge transfer, while inefficient unlabeled data learning causes confirmation bias.

Purpose of the Study:

• To propose a novel framework, bidirectional correlation maps domain adaptation (BCMDA), to address challenges in MiDSS.
• To enhance cross-domain learning and mitigate confirmation bias in semi-supervised medical image segmentation.

Main Methods:

• Knowledge transfer via virtual domain bridging (KTVDB) using bidirectional correlation maps to synthesize aligned virtual domains.
• Dual bidirectional CutMix for progressive knowledge transfer and prototypical alignment and pseudo label correction (PAPLC) to reduce confirmation bias.
• Utilizing learnable prototype cosine similarity classifiers for alignment and prototypical pseudo label correction for reliable pseudo-labels.

Main Results:

• BCMDA demonstrates superior performance on three public multi-domain datasets.
• The method shows excellent results even with a very limited number of labeled samples.
• The proposed framework effectively overcomes domain shift and confirmation bias issues in MiDSS.

Conclusions:

• The BCMDA framework offers a robust solution for mixed domain semi-supervised medical image segmentation.
• The approach significantly improves segmentation accuracy under domain shift and data scarcity.
• Code is publicly available for reproducibility and further research.