Mamba-Sea: A Mamba-Based Framework With Global-to-Local Sequence Augmentation for Generalizable Medical Image...

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Domain generalization (DG) is crucial for medical image segmentation to address distribution shifts between unseen target domains.
• Current DG methods primarily rely on Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs).
• Advanced state space models like Mamba show promise in medical image segmentation due to their ability to capture long-range dependencies with linear complexity.

Purpose of the Study:

• To explore the potential of Mamba architecture for domain generalization in medical image segmentation.
• To propose a novel Mamba-based framework, Mamba-Sea, to improve generalizability under domain shifts.
• To establish a new state-of-the-art in medical image segmentation using Mamba for DG.

Main Methods:

• Developed Mamba-Sea, a Mamba-based framework incorporating global-to-local sequence augmentation.
• Implemented a global augmentation mechanism to simulate appearance variations and reduce domain-specific learning.
• Introduced sequence-wise augmentation to perturb token styles by resampling style statistics related to domain shifts.

Main Results:

• Mamba-Sea demonstrates strong robustness to domain shifts in medical image segmentation.
• Achieved a Dice coefficient exceeding 90% on the Prostate dataset, surpassing the previous state-of-the-art (88.61%).
• This work is the first to explore Mamba generalization for medical image segmentation.

Conclusions:

• Mamba-Sea offers a promising Mamba-based architecture for robust medical image segmentation under domain shifts.
• The proposed global-to-local sequence augmentation effectively enhances model generalizability.
• The framework sets a new benchmark for domain generalization in medical imaging tasks.