Abstract
Deep learning models for medical image segmentation often suffer from "data hunger," requiring extensive annotated datasets to generalize effectively. While sequence models like Mamba excel at capturing long-range dependencies, they are prone to overfitting in low-data scenarios by memorizing irrelevant patterns. This paper introduces Differential-UMamba (Diff-UMamba), an architecture that integrates a specialized Noise Reduction Module (NRM) into a Mamba-based UNet. By explicitly modeling and subtracting noise in the latent space through a signal differencing strategy, Diff-UMamba achieves significant performance gains across multiple 3D medical imaging tasks, particularly when training data is scarce.
Introduction
Medical imaging has been transformed by deep learning, yet the scarcity of expert-labeled 3D volumes remains a major hurdle. Traditional Convolutional Neural Networks (CNNs) are robust but struggle to capture global context—the long-range relationships between distant anatomical structures. Newer architectures like Transformers and Mamba address this but are notoriously data-hungry. In limited-data regimes, these models often overfit, treating noise as clinical features. Diff-UMamba is designed to capture global context while filtering out these spurious features, making it ideal for clinical settings where data is limited.
Methods
The core innovation of Diff-UMamba is the Noise Reduction Module (NRM), situated at the network's bottleneck. It builds upon the UMamba-Bot architecture, combining the strengths of UNet and Mamba with a novel denoising mechanism.
The NRM operates via a "signal differencing" strategy:
- Feature Extraction: It aggregates outputs from every encoder layer ($e_1$ through $e_5$).
- Downsampling: These features are projected to a uniform spatial resolution using specialized DownSampling Blocks.
- Weighted Aggregation: Features are combined using learnable parameters ($\lambda_i$), allowing the model to prioritize specific layers for noise reduction.
- Noise Estimation: The aggregate signal is processed by a Mamba block to estimate the noise component ($m_2$).
- Subtraction: This estimated noise is subtracted from the primary bottleneck feature ($m_1$) to produce a refined representation ($\hat{m}$).
Datasets and Experimental Setup
The researchers validated Diff-UMamba across five diverse 3D medical imaging datasets, encompassing various modalities and segmentation targets.
To specifically test robustness under data scarcity, the BraTS-21 dataset was used to simulate varying levels of training data (from 16% to 80%). The model was also tested on a private Internal Dataset of non-small cell lung cancer (NSCLC) patients to ensure real-world applicability.
Key Insights
- Overfitting as Latent Noise: t-SNE visualizations revealed that models trained on limited data produce scattered feature clusters similar to those generated by pure Gaussian noise. The NRM effectively counteracts this "scattering."
- Adaptive Layer Prioritization: In low-data scenarios, the learnable $\lambda$ parameters for shallower encoder layers were significantly higher, suggesting that the NRM primarily targets noise introduced at early stages of feature extraction.
- Mamba's Inductive Bias: Mamba blocks were found to be superior to Transformers for this specific noise reduction task, likely due to their efficiency in modeling sequential dependencies in constrained environments.
Results
Diff-UMamba consistently outperformed state-of-the-art models, including nnUNetv2 and SwinUNETR:
- Public Benchmarks: Achieved a 1–3% gain in Intersection over Union (IoU) on MSD and AIIB23 datasets.
- Internal Validation: Showed a 4–5% improvement in segmenting gross tumor volume in cone-beam CT scans.
- Robustness: The model maintained high Dice scores even when artificial noise (e.g., Gaussian or Speckle) was introduced during inference, demonstrating the NRM's filtering capability.
- Efficiency: These improvements were achieved with only a 2% increase in total parameters, ensuring the model remains computationally viable.
Limitations and Future Work
While effective, the current NRM is restricted to the bottleneck to manage computational costs. Future research could explore integrating NRM into every encoder layer for finer-grained denoising. Additionally, testing the architecture on 2D imaging and non-medical domains could broaden its impact. The authors also suggest a "Dynamic NRM" that activates only when a learned noise threshold is met, potentially improving scalability for larger datasets.
Summary updated