Style mixup enhanced disentanglement learning for unsupervised domain adaptation in medical image segmentation

Zhuotong Cai ¹, Jingmin Xin ², Chenyu You ³

⁰National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an, Shaanxi, China; Department of Biomedical Engineering, Yale University, New Haven, CT, USA.

Medical Image Analysis +

|

January 7, 2025

View abstract on PubMed

Summary

This study introduces Style Mixup Enhanced Disentanglement Learning (SMEDL) for unsupervised domain adaptation in medical image segmentation. SMEDL improves model generalization and domain-invariant learning without image translation, outperforming existing methods.

Area Of Science

Medical Image Analysis
Computer Vision
Machine Learning

Background

Unsupervised domain adaptation (UDA) is crucial for medical image segmentation across different modalities.
Existing UDA methods often rely on image translation, which can compromise semantic consistency and domain-invariant representation.
There is a need for UDA approaches that enhance generalizability and domain invariance without explicit image translation.

Purpose Of The Study

To propose a novel method, Style Mixup Enhanced Disentanglement Learning (SMEDL), for UDA in medical image segmentation.
To improve model generalization and domain-invariant learning capabilities.
To overcome limitations of existing UDA methods dependent on image translation.

Main Methods

Employs disentangled style mixup to implicitly generate diverse style-mixed domains in the feature space.
Introduces pixel-wise consistency regularization for effective style-mixed domains and domain consistency.
Utilizes dual-level domain-invariant learning: intra-domain contrastive learning and inter-domain adversarial learning.

Main Results

Demonstrates superior performance on cardiac and brain medical image segmentation tasks.
Achieves improved domain generalization and domain-invariant learning.
Outperforms state-of-the-art methods in UDA medical image segmentation.

Conclusions

SMEDL offers an effective approach for UDA medical image segmentation by avoiding image translation.
The proposed method enhances model generalization and domain-invariant representation learning.
SMEDL shows significant potential for improving cross-modality medical image segmentation.

Keywords:

Disentanglement learning Medical image segmentation Style mixup Unsupervised domain adaptation