# U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation

**Authors:** Jun Ma, Feifei Li, Bo Wang  
**arXiv:** [2401.04722](https://arxiv.org/abs/2401.04722)  
**Date:** 2024-01-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2401.04722v1/

## Key points

- Popular architectures in biomedical image segmentation: and
- Limitations of and in handling long-range dependencies due to locality or computational complexity
- Introduction of a groundbreaking solution: , a versatile network designed for biomedical image segmentation
- Innovative design of : hybrid CNN-SSM block integrating local feature extraction with long-range dependency capturing
- Self-configuring mechanism of : automatic adaptation to diverse datasets without manual intervention
- Superior performance of over state-of-the-art CNN-based and Transformer-based networks in extensive experiments on various tasks
- Potential for efficient long-range dependency modeling in biomedical image analysis demonstrated by
- Spearheaded by researchers Jun Ma, Feifei Li, and Bo Wang
- Public accessibility of code, models, and data at https://wanglab.ai/u-mamba.html

## Summary

In the realm of biomedical image segmentation,  and  have stood out as popular architectures. However, their effectiveness is hindered by limitations in handling long-range dependencies due to inherent locality or computational complexity. To overcome this challenge, a groundbreaking solution has emerged in the form of , a versatile network designed specifically for biomedical image segmentation. Drawing inspiration from State Space Sequence Models (SSMs), renowned for their adeptness in managing long sequences,  introduces a hybrid CNN-SSM block. This innovative design seamlessly integrates the local feature extraction capabilities of convolutional layers with the long-range dependency capturing abilities of SSMs. One key advantage of  lies in its self-configuring mechanism, which enables automatic adaptation to diverse datasets without requiring manual intervention. Extensive experiments conducted on four distinct tasks - including 3D abdominal organ segmentation in CT and MR images, instrument segmentation in endoscopy images, and cell segmentation in microscopy images - have demonstrated the superior performance of  over state-of-the-art CNN-based and Transformer-based segmentation networks across all tasks. This groundbreaking advancement not only showcases the potential for efficient long-range dependency modeling in biomedical image analysis but also paves the way for new avenues of research and development in this critical field. Researchers Jun Ma, Feifei Li, and Bo Wang have spearheaded this transformative work on . The code, models, and data associated with this innovative network are publicly accessible at https://wanglab.ai/u-mamba.html.