Vision Mamba Empowered by Dynamic Domain Generalization for Cross-Modality Medical Segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning excels in medical image segmentation but struggles with domain shift, especially in cross-modal tasks.
• Existing unsupervised domain adaptation (UDA) and domain generalization (DG) methods have limitations, including data acquisition challenges for UDA and overlooking anatomical priors in DG.

Purpose of the Study:

• To propose a novel cross-modal medical image segmentation framework.
• To enhance model generalization across different medical imaging modalities.
• To address limitations of current UDA and DG techniques in medical imaging.

Main Methods:

• Integration of the Vision Mamba model with dynamic domain generalization.
• Utilizing bidirectional state-space sequence modeling, Bezier curve-style enhancement, and dual-normalization for feature alignment and fusion.
• Introduction of the VEBlock module combining Mamba's dynamic sequence modeling with non-local attention for cross-modal dependencies.

Main Results:

• Significant improvements in cross-modal segmentation performance on BraTS 2018 and cardiac datasets.
• Achieved an average Dice score of 56.22% in T2→T1 tasks, outperforming baselines by 1.78%.
• Reduced Hausdorff distance for tumor boundaries to 13.26 mm and optimized cardiac CT→MRI Hausdorff distance to 27.34 mm.

Conclusions:

• The proposed framework demonstrates strong generalization capabilities for cross-modal medical image segmentation.
• Effectively addresses challenges related to domain shift and anatomical priors in medical imaging.
• Offers a promising approach for improving segmentation accuracy and reliability across diverse medical imaging modalities.