A-DualUD: Source-free unsupervised domain adaptation via anatomical anchor alignment and dual-path uncertainty denoising for cross-modality medical image segmentation.

BACKGROUND AND OBJECTIVE

Automated medical image segmentation across different imaging modalities, such as MRI and CT scans, plays a crucial role in improving diagnostic accuracy, treatment planning, and surgical navigation. However, significant domain discrepancies across different modalities, stemming from various acquisition protocols and physical principles, pose a significant challenge for automated segmentation algorithms. Unsupervised domain adaptation (UDA) has gained considerable attention for addressing challenges in cross-modality medical image segmentation from different medical instruments, particularly in scenarios where target domain labels are unavailable. However, traditional UDA methods require simultaneous access to source domain data, which limits their practical applicability in medical scenarios where source data is often inaccessible due to privacy concerns and storage restrictions.

METHODS

In this work, we propose a novel source free unsupervised domain adaptation (SFUDA) framework, which introduces anatomical anchor alignment (A) and dual-path uncertainty denoising (DualUD) to facilitate knowledge transfer from a pre-trained source model to an unlabeled target domain, without requiring access to source data. In the A stage, we extract the source domain anatomical anchors that represent characteristics of each target class and align them with the target domain features through a bidirectional constraint approach, thereby reducing feature-level distribution discrepancies between the source and target domains. Additionally, we introduce DualUD stage to provide sufficient and reliable supervision for domain adaptation.

RESULTS

Extensive experiments conducted on cross-modality multi-organ segmentation tasks, using the abdominal and cardiac datasets, demonstrate the state-of-the-art performance of our proposed approach.

CONCLUSION

Anatomical anchor alignment and dual-path uncertainty denoising effectively address the critical challenges of domain discrepancies and privacy preservation, offering a robust and effective solution for achieving SFUDA. The source code is publicly available at: https://github.com/derekshiii/A3-DualUD.