Differential-UMamba: Rethinking tumor segmentation under limited data scenarios

Dhruv Jain, Romain Modzelewski, Romain Hérault, Clément Chatelain, Eva Torfeh, Sébastien Thureau

TL;DR

Differential-UMamba is a new architecture designed to solve the "data hunger" problem in medical image segmentation. By integrating a specialized Noise Reduction Module (NRM) into a Mamba-based UNet, the model learns to filter out spurious features and noise, leading to significant performance gains in scenarios where training data is scarce.

What Problem Does This Paper Address?

Deep learning has transformed medical imaging, but it has a major Achilles' heel: it requires massive amounts of annotated data. In the medical world, expert-labeled 3D volumes are rare and expensive to produce.

While traditional Convolutional Neural Networks (CNNs) are robust, they struggle to capture "long-range dependencies"—the relationship between distant parts of an organ or tumor. Newer sequence models like Transformers and Mamba excel at this, but they are notoriously prone to overfitting. When data is limited, these models often "memorize" noise and irrelevant patterns instead of learning the actual clinical features. The authors of this paper set out to create a model that captures global context without falling into the trap of overfitting on small datasets.

Methods

The authors introduce Diff-UMamba, which builds upon the UMamba-Bot architecture (a hybrid of UNet and Mamba). The core innovation is the Noise Reduction Module (NRM) located in the network's bottleneck.

An overview of the Diff-UMamba architecture, highlighting the Noise Reduction Module (NRM) and the signal differencing strategy.

The NRM works through a "signal differencing" strategy:

  1. Feature Extraction: It takes outputs from every encoder layer ($e_1$ through $e_5$).
  2. Downsampling: These features are projected to a uniform spatial resolution.
  3. Weighted Aggregation: The features are combined using learnable parameters ($\lambda_i$).
  4. Noise Estimation: This aggregate is passed through a dedicated Mamba block to estimate the "noise" component ($m_2$).
  5. Subtraction: This estimated noise is subtracted from the primary bottleneck feature ($m_1$).

By explicitly modeling and subtracting noise in the latent space, the model is forced to focus on task-relevant representations.

Datasets

The researchers evaluated their model across a diverse range of 3D medical imaging tasks:

Summary of the five datasets used in the study, ranging from lung and pancreas CTs to brain MRIs.

Notably, they used the BraTS-21 dataset to simulate different levels of data scarcity by training on varying proportions (16% to 80%) of the total data. They also validated the approach on a private Internal Dataset of non-small cell lung cancer (NSCLC) patients.

Key Insights and Ideas

  • Overfitting as Noise: The authors used t-SNE visualizations to show that the latent space of a model trained on limited data looks remarkably similar to a model being fed actual Gaussian noise. Both result in "scattered" and incoherent feature clusters.
  • The "Aha" Moment for $\lambda$: In low-data regimes, the learnable $\lambda$ parameters for the earlier (shallower) encoder layers were significantly higher. This suggests that noise predominantly enters the system at the initial stages of feature extraction, and the NRM successfully learns to prioritize cleaning those layers.
  • Diminishing Returns with More Data: As the training set size increases, the $\lambda$ values naturally decrease. The model "realizes" it doesn't need to rely as heavily on the NRM when it has enough data to learn robust features on its own.
  • Mamba's Inductive Bias: The study found that Mamba blocks are better suited for integrating this noise reduction strategy than Transformers, likely due to their superior ability to model sequential patterns in data-constrained environments.

Key Results

Diff-UMamba consistently outperformed state-of-the-art models like nnUNetv2, SwinUNETR, and the baseline UMamba:

  • Public Datasets: Achieved a 1–3% gain in Intersection over Union (IoU) across the MSD and AIIB23 datasets.
  • Internal Dataset: Showed a substantial 4–5% improvement in segmenting gross tumor volume in cone-beam CT scans.
  • Robustness: When artificial noise (Gaussian, Speckle, etc.) was injected during inference, Diff-UMamba maintained much higher Dice scores than the baseline, proving the NRM's effectiveness as a filter.
  • Efficiency: These gains were achieved with only a 2% increase in total parameters, making it a highly efficient upgrade.

Limitations and Future Work

While the results are promising, the authors acknowledge several areas for growth:

  • Computational Cost: Currently, the NRM is only in the bottleneck. Integrating it into every encoder layer could improve performance further but would be computationally expensive due to high-resolution feature maps.
  • Scope: The study focused exclusively on 3D medical segmentation. Its effectiveness in 2D imaging or non-medical domains (like natural image processing) remains untested.
  • Dynamic Activation: Future iterations could include a "Dynamic NRM" that only activates when a learned noise threshold is exceeded, potentially making the model even more scalable for large-scale datasets.

Summary updated

BibTeX
@article{jain2026differential,
  author = {Dhruv Jain and Romain Modzelewski and Romain Hérault and Clément Chatelain and Eva Torfeh and Sébastien Thureau},
  title = {{Differential-UMamba: Rethinking tumor segmentation under limited data scenarios}},
  journal = {Biomedical Signal Processing and Control},
  year = {2026},
  doi = {10.1016/j.bspc.2026.110163},
}
Figure