DuetMatch: Harmonizing semi-supervised brain MRI segmentation via decoupled branch optimization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Limited annotated data in medical imaging necessitates advanced learning techniques.
• Semi-supervised learning (SSL) offers a solution by leveraging imperfectly labeled data.
• Teacher-student frameworks show promise but face convergence challenges in complex scenarios.

Purpose of the Study:

• To introduce DuetMatch, a novel dual-branch SSL framework for robust medical image segmentation.
• To address stability and convergence issues in existing teacher-student models.
• To enhance performance in scenarios with limited annotated medical imaging data.

Main Methods:

• Proposed DuetMatch, a dual-branch framework with asynchronous optimization (encoder/decoder branches optimize alternately).
• Introduced Decoupled Dropout Perturbation for improved consistency under noisy conditions.
• Implemented Pairwise CutMix Cross-Guidance and Consistency Matching to enhance diversity and mitigate confirmation bias.

Main Results:

• DuetMatch demonstrated superior performance compared to state-of-the-art methods on benchmark brain MRI segmentation datasets (ISLES2022, BraTS).
• The framework showed consistent effectiveness and robustness across various SSL segmentation tasks.
• Asynchronous optimization and proposed regularization techniques proved beneficial for challenging segmentation scenarios.

Conclusions:

• DuetMatch offers a robust and effective solution for semi-supervised medical image segmentation.
• The novel techniques significantly improve model stability, diversity, and accuracy.
• The framework shows great potential for applications with limited annotated medical imaging data.