Abstract
Mamba-Sea is a novel framework designed to enhance the domain generalization of medical image segmentation models using State Space Models (SSMs). While Mamba architectures offer linear computational complexity and global receptive fields, they often overfit to domain-specific styles. Mamba-Sea addresses this through a "Global-to-Local" sequence augmentation strategy, comprising Global Appearance Variation Augmentation (GVA) and Local Sequence-wise Style Transformation Augmentation (LSA). Evaluated across diverse medical imaging datasets, Mamba-Sea achieves state-of-the-art performance, notably surpassing the 90% Dice coefficient threshold on prostate MRI segmentation, while maintaining high efficiency during inference.
Introduction
The deployment of medical AI in clinical settings is frequently hindered by "domain shift"—the performance degradation that occurs when a model trained on data from one source (e.g., a specific scanner or protocol) is applied to data from another. Traditional solutions like Convolutional Neural Networks (CNNs) often lack global context, while Vision Transformers (ViTs) face quadratic scaling issues with high-resolution images.
Recently, Mamba (Structured State Space Models) has emerged as a promising alternative, providing global receptive fields with linear complexity. However, Mamba’s input-dependent mechanisms tend to "memorize" domain-specific styles, limiting its generalizability. Mamba-Sea is proposed to mitigate this memorization, forcing the model to learn robust, domain-invariant anatomical features.
Methodology
The Mamba-Sea framework utilizes a VM-Unet backbone and introduces a hierarchical augmentation strategy to suppress domain-specific noise. The core architecture is built around the VSSDG block, which integrates these augmentations into the training pipeline.
Global Appearance Variation Augmentation (GVA)
This module employs a lightweight network to simulate broad variations in lighting and contrast typical of different clinical sites. A gating mechanism ensures that augmentations are only applied when beneficial, preventing the degradation of already well-illuminated scans.
Local Sequence-wise Style Transformation Augmentation (LSA)
Since Mamba processes images as sequences of tokens, LSA targets the "style" at the sequence level. By modeling feature statistics (mean and standard deviation) as Gaussian distributions and resampling them for random continuous sub-sequences, LSA disrupts the model's tendency to rely on local textures or domain-specific styles.
Semantic Consistency Training
To ensure that these augmentations do not alter the underlying anatomy, the model is trained with a consistency loss. This requires the model to produce identical segmentation outputs for both original and augmented image pairs, reinforcing the learning of domain-invariant features.
Experimental Results
The framework was rigorously tested on three distinct medical imaging tasks to prove its robustness and scalability:
- Fundus Imaging: Optic cup/disc segmentation across four public datasets (Drishti-GS, RIM-ONE-R3, REFUGE). Mamba-Sea achieved an average Dice coefficient of 89.68%, a significant 2.78% improvement over the baseline VM-Unet.
- Prostate MRI: Segmentation of T2-weighted images from six clinical sources. Mamba-Sea is the first framework to surpass the 90% Dice coefficient barrier (90.02%) on this challenging task, significantly outperforming previous state-of-the-art methods.
- Skin Lesion Segmentation: Demonstrated scalability on a large-scale dataset, outperforming specialized foundation model adaptations like Medical SAM.
Crucially, the GVA and LSA modules are only active during training. At inference, Mamba-Sea operates as a standard, efficient Mamba model, providing state-of-the-art generalization with no additional computational overhead. Training overhead is also minimal, increasing iteration time by only 0.004 seconds.
Conclusion
Mamba-Sea represents a significant step forward in generalizable medical image segmentation. By combining the efficiency of State Space Models with a targeted global-to-local augmentation strategy, it effectively overcomes the domain-memorization pitfalls of standard Mamba architectures. Future work will explore class-aware augmentation strategies and the potential of Mamba-Sea as a multi-modal foundation model for medical imaging.
Summary updated