TL;DR
Mamba-Sea is a novel framework that brings the efficiency of Mamba (State Space Models) to the challenge of domain generalization in medical image segmentation. By introducing a "Global-to-Local" sequence augmentation strategy, the authors successfully suppress domain-specific noise, allowing the model to achieve state-of-the-art performance on unseen clinical data while maintaining linear computational complexity.
What Problem Does This Paper Address?
In clinical practice, medical AI models often encounter "domain shift." A model trained on MRI scans from one hospital (Source Domain) might fail miserably when deployed at another hospital using a different scanner or imaging protocol (Target Domain).
Existing solutions usually rely on Convolutional Neural Networks (CNNs), which have limited receptive fields and tend to overfit to local textures, or Vision Transformers (ViTs), which capture global context but suffer from quadratic computational costs as image resolution increases. Recently, Mamba (a Structured State Space Model) has emerged as a "best of both worlds" architecture—offering global receptive fields with linear complexity. However, Mamba’s internal mechanisms (specifically its input-dependent matrices) tend to "memorize" domain-specific styles during training. The authors of Mamba-Sea set out to fix this "memorization" problem to make Mamba truly generalizable.
Methods
The core of the paper is the Mamba-Sea framework, which uses a VM-Unet backbone and introduces three key components to force the model to learn domain-invariant features:
- Global Appearance Variation Augmentation (GVA): This module uses a lightweight, learnable network to simulate broad variations in image appearance (like lighting or contrast changes) across different clinical sites. It includes a clever gating mechanism that only applies augmentation to images that need it, 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" of these sequences. It models feature statistics (mean and standard deviation) as Gaussian distributions and resamples them for random continuous sub-sequences. This "perturbs" the local style, forcing the Mamba blocks to focus on the underlying anatomy rather than the specific "look" of the data.
- Semantic Consistency Training: The model is trained to produce identical segmentation masks for both the original image and its augmented versions. This consistency loss ensures that the perturbations introduced by GVA and LSA don't confuse the model's understanding of the anatomy.
Datasets
The authors evaluated Mamba-Sea on three diverse tasks to prove its robustness:
- Fundus Dataset: Optic cup/disc segmentation using 789 training and 281 test cases from four different public sources (Drishti-GS, RIM-ONE-R3, REFUGE).
- Prostate Dataset: T2-weighted MRI segmentation comprising 116 images from six different clinical sources.
- Skin Lesion Dataset: A large-scale dataset (S1 → T1) featuring 3,694 training images and 200 testing images, used to test the model's scalability.
Key Insights and Ideas
- Mamba's "Memory" is a Double-Edged Sword: While Mamba’s ability to compress sequence information into hidden states is powerful, it naturally accumulates domain-specific "noise." LSA is specifically designed to disrupt this accumulation at the sequence level.
- Efficiency at Inference: A major advantage of Mamba-Sea is that the GVA and LSA modules are only used during training. At inference time, they are discarded, meaning the model provides state-of-the-art generalization with no extra computational cost compared to a standard Mamba model.
- Surpassing the "90% Barrier": The framework is the first to surpass a 90% Dice coefficient on the challenging Prostate MRI dataset, a significant milestone in the field of domain generalization.
Key Results
Mamba-Sea consistently outperformed CNN, ViT, and even recent "Foundation Model" adaptations (like Medical SAM):
- Prostate MRI: Achieved a 90.02% Dice coefficient, significantly beating the previous state-of-the-art (88.61%).
- Fundus Imaging: Reached an average Dice of 89.68%, a 2.78% improvement over the baseline VM-Unet.
- Skin Lesions: Achieved 93.11% Dice, outperforming specialized SAM-based methods like ESP-MedSAM.
- Computational Cost: The training overhead was negligible (increasing iteration time by only 0.004 seconds), and it maintained a competitive parameter count (~27.43M).
Limitations and Future Work
While Mamba-Sea shows remarkable robustness, the authors acknowledge a few areas for growth:
- Persistent Domain Info: Some domain-specific information may still persist in complex real-world images that the current augmentations don't fully capture.
- Class-Aware Augmentation: The current LSA/GVA strategies are class-agnostic. Future versions could benefit from "class-aware" strategies that perturb different anatomical structures differently.
- Scaling Up: The authors suggest exploring Mamba-Sea as a foundation model by training on even larger, multi-modal datasets to further push the boundaries of medical AI generalization.
Summary updated