Related Experiment Video
Updated: May 6, 2026
A Pipeline for 3D Multimodality Image Integration and Computer-assisted Planning in Epilepsy Surgery
Published on: May 20, 2016
MambaMIM: Pre-training Mamba with state space token interpolation and its application to medical image segmentation
Fenghe Tang1, Bingkun Nian2, Yingtai Li1
1School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, 230026, PR China; Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou, Jiangsu, 215123, PR China; Center for Medical Imaging, Robotics, and Analytic Computing & LEarning (MIRACLE), Suzhou Institute for Advanced Research, USTC, Suzhou, Jiangsu, 215123, PR China.
MambaMIM, a new masked image modeling framework, enhances Mamba's capabilities in 3D medical imaging by leveraging causal state space properties. This approach achieves state-of-the-art performance in medical image segmentation tasks.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computer Vision
Background:
- State space models like Mamba show promise for long-sequence modeling in 3D medical imaging.
- Existing generative self-supervised methods do not fully exploit Mamba's potential for long-range dependencies due to overlooking causal properties in masked modeling.
Purpose of the Study:
- To introduce MambaMIM, a general-purpose pre-training framework designed to unlock Mamba's capabilities in medical image analysis.
- To enhance Mamba's multi-scale and long-range representation abilities for medical imaging tasks.
Main Methods:
- Developed MambaMIM, a masked image modeling approach utilizing a novel TOKen-Interpolation (TOKI) strategy for selective state space sequences.
- Implemented a bottom-up 3D hybrid masking strategy for consistent masking across architectures.
- Pre-trained MambaMIM on a large dataset of 6.8K CT scans.
Main Results:
- MambaMIM demonstrated significant advancements in medical image pre-training.
- Achieved state-of-the-art segmentation performance when applied to a hybrid MedNeXt and Vision Mamba architecture.
- Validated feasibility and advancement of Mamba for medical image pre-training across eight public benchmarks.
Conclusions:
- MambaMIM effectively addresses limitations in current generative self-supervised learning for Mamba in medical imaging.
- The proposed framework enhances Mamba's performance in long-sequence medical visual tasks, particularly in segmentation.
- The study highlights Mamba's potential as a powerful tool for future medical image analysis research.
