U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation: U-Mamba Hybridizes CNN and State Space Models to Achieve Long-range Dependency Modeling in Biomedical Image Segmentation, Surpassing CNNs and Transformers
U-Mamba proposes a hybrid CNN-SSM module that combines convolutional local feature extraction with the long-range dependency capture capability of state space sequence models, outperforming existing CNN and Transformer networks on four segmentation tasks: CT/MR abdominal organ, endoscopic instrument, and microscopy cell segmentation.
U-Mamba introduces state space models (SSM) into the U-Net architecture, designing a hybrid CNN-SSM module that effectively addresses the long-range dependency challenges of CNN locality and Transformer computational complexity, achieving state-of-the-art performance on multiple biomedical segmentation benchmarks.
The core of U-Mamba is the hybrid CNN-SSM module, which integrates convolutional layers with Mamba (an SSM-based sequence model) in parallel or serial fashion, capturing global context while preserving local details. The network adopts a U-Net-style encoder-decoder structure with a self-configuration mechanism that automatically adapts to different datasets. Experiments cover four tasks: CT and MR abdominal multi-organ segmentation (Dice improvement of 1-3%), endoscopic instrument segmentation, and microscopy cell segmentation, all outperforming baselines such as nnU-Net and SwinUNETR. Code and models are open-sourced.
U-Mamba provides an efficient long-range dependency modeling solution for medical image analysis, potentially replacing the dominant role of Transformers in medical AI, reducing computational costs, and advancing applications such as real-time diagnosis and surgical navigation.
It is recommended that medical AI companies evaluate U-Mamba as an alternative to existing segmentation models, prioritize integration testing in CT/MR organ segmentation and surgical robot vision systems, and leverage its self-configuration feature to reduce data adaptation costs.
Key validation signals include: generalization testing on larger multi-center datasets, fair comparison with nnU-Net, inference speed evaluation on edge devices, and regulatory approval progress for clinical deployment.