A New Bidirectional Unsupervised Domain Adaptation Segmentation Framework

Munan Ning, Cheng Bian, Dong Wei, Chenglang Yuan, Yaohua Wang, Yang Guo, Kai Ma, Yefeng Zheng

Abstract

Domain shift happens in cross-domain scenarios commonly because of the wide gaps between different domains: when applying a deep learning model well-trained in one domain to another target domain, the model usually performs poorly. To tackle this problem, unsupervised domain adaptation (UDA) techniques are proposed to bridge the gap between different domains, for the purpose of improving model performance without annotation in the target domain. Particularly, UDA has a great value for multimodal medical image analysis, where annotation difficulty is a practical concern. However, most existing UDA methods can only achieve satisfactory improvements in one adaptation direction (e.g., MRI to CT), but often perform poorly in the other (CT to MRI), limiting their practical usage. In this paper, we propose a bidirectional UDA (BiUDA) framework based on disentangled representation learning for equally competent two-way UDA performances. This framework employs a unified domain-aware pattern encoder which not only can adaptively encode images in different domains through a domain controller, but also improve model efficiency by eliminating redundant parameters. Furthermore, to avoid distortion of contents and patterns of input images during the adaptation process, a content-pattern consistency loss is introduced. Additionally, for better UDA segmentation performance, a label consistency strategy is proposed to provide extra supervision by recomposing target-domain-styled images and corresponding source-domain annotations. Comparison experiments and ablation studies conducted on two public datasets demonstrate the superiority of our BiUDA framework to current state-of-the-art UDA methods and the effectiveness of its novel designs. By successfully addressing two-way adaptations, our BiUDA framework offers a flexible solution of UDA techniques to the real-world scenario.

Introduction

Deep learning models for medical imaging are notoriously sensitive to domain shift. A model trained to segment a heart in an MRI scan often fails when applied to a CT scan because the "appearance" (intensity, noise, contrast) differs wildly, even if the underlying anatomy (the "content") is the same.

Unsupervised Domain Adaptation (UDA) aims to mitigate this by adapting a model from a labeled source domain to an unlabeled target domain. However, current state-of-the-art (SOTA) methods often suffer from a "one-way street" limitation: they may perform well for MRI $\to$ CT but poorly for CT $\to$ MRI. This paper introduces BiUDA, a novel framework designed to achieve equally competent bidirectional performance, addressing the "domain drop" phenomenon that limits practical clinical usage.

Methodology

The BiUDA framework is built on the foundation of Disentangled Representation Learning (DRPL). It decomposes an image into two distinct components: a domain-invariant Content Code ($c$) representing anatomical structure and a domain-specific Pattern Code ($p$) representing modality-specific appearance.

The BiUDA framework architecture, featuring a domain-aware pattern encoder and disentangled representation learning.

Key Innovations

  • Domain-Aware Pattern Encoder ($E_p$): Unlike previous works that use separate encoders for each domain, BiUDA employs a single, unified encoder. A "domain controller" input tells the encoder which modality it is processing, reducing parameters and forcing the model to explicitly learn inter-domain differences.
  • Content-Pattern Consistency Loss ($\mathcal{L}^{cpc}$): To prevent the model from distorting anatomy during translation, this loss ensures that decomposed and then recomposed codes remain consistent with the originals.
  • Label Consistency Strategy: The framework creates pseudo-training data by applying the pattern of the target domain to the anatomical content of a labeled source image. This allows the segmenter to learn target-domain appearances while utilizing ground-truth source annotations.

Experimental Results

The authors evaluated BiUDA on two challenging multi-modality datasets: MMWHS (Whole Heart Segmentation) and MMAS (Abdominal Segmentation). The framework was compared against SOTA methods including SIFA, ACE, and DISE.

Comparison of performance drop across different UDA methods for forward and backward adaptation directions.

The primary metric used was the performance drop—the gap between the UDA model and a "gold standard" model trained directly on target labels.

  • MMWHS Dataset: BiUDA achieved a mean Dice performance drop of only 8.49%, significantly lower than ACE (13.47%) and SIFA (21.84%).
  • MMAS Dataset: BiUDA led with a mean Dice drop of 4.39%, outperforming ACE (5.27%) and SIFA (7.52%).
  • Bidirectional Stability: As shown in Figure 2, the gap between "Forward" and "Backward" performance is much narrower in BiUDA than in competing frameworks, proving its reliability regardless of which modality is labeled.

Conclusion

BiUDA successfully addresses the asymmetry in current UDA methods by leveraging a unified domain-aware encoder and strict content-pattern consistency. By treating anatomy as a stable anchor and modality as a swappable style, the framework provides a robust solution for real-world medical scenarios where labels may be scarce for any given modality. Future work includes extending the framework to 3D volumes and exploring its generalization to other modalities like PET or Ultrasound.

Summary updated

Abstract

Domain shift happens in cross-domain scenarios commonly because of the wide gaps between different domains: when applying a deep learning model well-trained in one domain to another target domain, the model usually performs poorly. To tackle this problem, unsupervised domain adaptation (UDA) techniques are proposed to bridge the gap between different domains, for the purpose of improving model performance without annotation in the target domain. Particularly, UDA has a great value for multimodal medical image analysis, where annotation difficulty is a practical concern. However, most existing UDA methods can only achieve satisfactory improvements in one adaptation direction (e.g., MRI to CT), but often perform poorly in the other (CT to MRI), limiting their practical usage. In this paper, we propose a bidirectional UDA (BiUDA) framework based on disentangled representation learning for equally competent two-way UDA performances. This framework employs a unified domain-aware pattern encoder which not only can adaptively encode images in different domains through a domain controller, but also improve model efficiency by eliminating redundant parameters. Furthermore, to avoid distortion of contents and patterns of input images during the adaptation process, a content-pattern consistency loss is introduced. Additionally, for better UDA segmentation performance, a label consistency strategy is proposed to provide extra supervision by recomposing target-domain-styled images and corresponding source-domain annotations. Comparison experiments and ablation studies conducted on two public datasets demonstrate the superiority of our BiUDA framework to current state-of-the-art UDA methods and the effectiveness of its novel designs. By successfully addressing two-way adaptations, our BiUDA framework offers a flexible solution of UDA techniques to the real-world scenario.
BibTeX
@inproceedings{ning2021new,
  author = {Munan Ning and Cheng Bian and Dong Wei and Chenglang Yuan and Yaohua Wang and Yang Guo and Kai Ma and Yefeng Zheng},
  title = {{A New Bidirectional Unsupervised Domain Adaptation Segmentation Framework}},
  booktitle = {IPMI},
  year = {2021},
  doi = {10.1007/978-3-030-78191-0\_38},
}
Figure