• We propose U-Mamba, A general-purpose segmentation network for 2D and 3D biomedical images.
  • U-Mamba combines the advantage of convolutional layers and state space models, which can simultaneously capture local features and aggregate long-range dependencies.
  • U-Mamba enjoys a self-configuring mechanism, allowing it to automatically adapt to various datasets without manual intervention.
  • Extensive experiments demonstrate U-Mamba achieves superior performance than both CNN- and Transformer-based networks on four diverse tasks, including the 3D abdominal organ segmentation in CT and MR images, instrument segmentation in endoscopy images, and cell segmentation in microscopy images.
  • The results suggest that U-Mamba is a promising candidate for serving the next-generation biomedical image segmentation network backbone.


Overview of the U-Mamba (Enc) architecture. a. U-Mamba building block contains two successive Residual blocks followed by a Mamba block to enhance long-range dependencies. b. The whole architecture of U-Mamba, which is a typical encoder-decoder framework with U-Mamba blocks in the encoder, Residual blocks in the decoder, together with skip connections.

Experiments and Results

Dataset Overview

We evaluate U-Mamba on four publicly available datasets, covering four modalities and various targets. U-Mamba inherits the self-configuring feature from nnU-Net, which can automatically generate networks for different datasets.

Table 1. Dataset information.
Dataset Dimension #Training Image #Testing Image #Targets
Abdomen CT 3D 50 (4794 slices) 50 (10894 slices) 13
Abdomen MRI 3D 60 (5615 slices) 50 (3357 slices) 13
Endoscopy Images 2D 1800 1200 7
Microscopy Images 2D 1000 101 2
Table 2. U-Mamba configurations for each dataset.
Configurations Patch size Batch size # Stages # Pooling per axis
Abdomen CT (40, 224, 192) 2 6 (3, 3, 5)
3D Abdomen MR (48, 160, 224) 2 6 (3, 5, 5)
2D Abdomen MR (320, 320) 30 7 (6, 6)
Endoscopy (384, 640) 13 7 (6, 6)
Microscopy (512, 512) 12 8 (7, 7)


Visualized results demonstrate that U-Mamba can handle complex targets with fewer segmentation outliers.

Table 3. Results summary of 3D organ segmentation on abdomen CT and MRI datasets. U-Mamba_Bot: only use the U-Mmaba block in the bottleneck. UMamba_ Enc: all encoder blocks are U-Mmaba blocks.
Visualized segmentation examples of abdominal organ segmentation in CT (1st and 2nd rows) and MRI scans (3rd and 4th rows).
Table 4. Results summary of 2D segmentation tasks: organ segmentation in abdomen MRI scans, instruments segmentation in endoscopy images, and cell segmentation in microscopy images.
Visualized segmentation examples of abdominal organ segmentation MRI scans (1st row), cell segmentation in microscopy images (2nd row), and instruments segmentation in endoscopy images (3rd row).


@article{U-Mamba, title={U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation}, author={Ma, Jun and Li, Feifei and Wang, Bo}, journal={arXiv preprint arXiv:2401.04722}, year={2024} }

This website template is borrowed from here. Thank you!