← ML Research Wiki / 2401.13560

SegMamba: Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation

Zhaohu Xing The Hong Kong University of Science and Technology (Guangzhou, Tian Ye The Hong Kong University of Science and Technology (Guangzhou, Yijun Yang The Hong Kong University of Science and Technology (Guangzhou, Guang Liu Beijing Academy of Artificial Intelligence, Lei Zhu [email protected] The Hong Kong University of Science and Technology (Guangzhou The Hong Kong University of Science and Technology (2024)

Paper Information
arXiv ID
2401.13560
Venue
International Conference on Medical Image Computing and Computer-Assisted Intervention
Domain
Medical image analysis
SOTA Claim
Yes
Code
Available
Reproducibility
7/10
Contents

Abstract

The Transformer architecture has demonstrated remarkable results in 3D medical image segmentation due to its capability of modeling global relationships.However, it poses a significant computational burden when processing high-dimensional medical images.Mamba, as a State Space Model (SSM), has recently emerged as a notable approach for modeling long-range dependencies in sequential data, and has excelled in the field of natural language processing with its remarkable memory efficiency and computational speed.Inspired by this, we devise SegMamba, a novel 3D medical image Segmentation Mamba model, to effectively capture long-range dependencies within whole-volume features at every scale.Our SegMamba outperforms Transformer-based methods in wholevolume feature modeling, maintaining high efficiency even at a resolution of 64 × 64 × 64, where the sequential length is approximately 260k.Moreover, we collect and annotate a novel large-scale dataset (named CRC-500) to facilitate benchmarking evaluation in 3D colorectal cancer (CRC) segmentation.Experimental results on our CRC-500 and two public benchmark datasets further demonstrate the effectiveness and universality of our method.The code for SegMamba is publicly available at: https://github.com/ge-xing/SegMamba.

Summary

This paper presents SegMamba, a novel approach for 3D medical image segmentation that efficiently captures long-range dependencies using a State Space Model (SSM) called Mamba. The method combines a tri-orientated Mamba (ToM) module with a gated spatial convolution (GSC) module and a feature-level uncertainty estimation (FUE) module. Extensive experiments demonstrate its superior performance on the novel CRC-500 dataset, which contains 500 annotated 3D colorectal cancer scans, as well as on two public benchmark datasets: BraTS2023 and AIIB2023. Results show that SegMamba outperforms existing methods in segmentation accuracy while maintaining high efficiency.

Methods

This paper employs the following methods:

  • Mamba
  • Tri-orientated Mamba (ToM)
  • Gated Spatial Convolution (GSC)
  • Feature-level Uncertainty Estimation (FUE)

Models Used

  • SegMamba
  • U-Mamba
  • UNETR
  • SwinUNETR
  • SwinUNETR-V2
  • SegresNet
  • UX-Net
  • MedNeXt

Datasets

The following datasets were used in this research:

  • CRC-500
  • BraTS2023
  • AIIB2023

Evaluation Metrics

  • Dice
  • Hausdorff Distance (HD95)
  • Intersection over Union (IoU)
  • Detected Length Ratio (DLR)
  • Detected Branch Ratio (DBR)

Results

  • SegMamba outperforms Transformer-based methods in whole-volume feature modeling
  • Achieved highest Dice scores of 93.61%, 92.65%, and 87.71% on the BraTS2023 dataset
  • Obtained best Dice and HD95 scores of 48.46% and 28.52 on the CRC-500 dataset
  • Achieved highest IoU, DLR, and DBR scores on AIIB2023 dataset

Technical Requirements

  • Number of GPUs: 4
  • GPU Type: NVIDIA A100

Keywords

SegMamba long-range dependencies 3D medical image segmentation state space model

Papers Using Similar Methods

External Resources

Funding: Not specified
References: 33
Influential Citations: 6