Adaptive Hierarchical Dual Consistency for Semi-Supervised Left Atrium Segmentation on Cross-Domain Data.

Area of Science:

• Medical image analysis
• Machine learning
• Computer-aided diagnosis

Background:

• Semi-supervised learning is crucial for left atrium (LA) segmentation with limited labeled data.
• Generalizing semi-supervised models to cross-domain data enhances robustness but is hindered by distribution differences and sample mismatch.
• Existing methods struggle with domain shift in medical image segmentation tasks.

Purpose of the Study:

• To propose an Adaptive Hierarchical Dual Consistency (AHDC) method for cross-domain semi-supervised left atrium segmentation.
• To address distribution differences and sample mismatch between different medical imaging domains.
• To improve the generalization capability and robustness of segmentation models.

Main Methods:

• Developed Adaptive Hierarchical Dual Consistency (AHDC) comprising Bidirectional Adversarial Inference (BAI) and Hierarchical Dual Consistency (HDC) modules.
• BAI module uses adversarial learning to align distributions and match samples across domains.
• HDC module employs a hierarchical dual learning paradigm for intra-domain and inter-domain consistency constraints.

Main Results:

• AHDC demonstrated superior performance on diverse 3D LGE-CMR and 3D CT datasets.
• Achieved higher segmentation accuracy compared to state-of-the-art methods in cross-domain settings.
• Validated the effectiveness of AHDC in handling distribution shifts and sample mismatches.

Conclusions:

• The proposed AHDC method effectively enhances cross-domain semi-supervised segmentation for left atrium.
• AHDC offers a robust solution for generalizing segmentation models to unseen data distributions.
• This approach significantly improves segmentation accuracy and model robustness in challenging cross-domain scenarios.