Related Experiment Video
Updated: May 24, 2025
12:30
A Comprehensive Protocol for Manual Segmentation of the Medial Temporal Lobe Structures
Published on: July 2, 2014
20.2K
MambaSAM: A Visual Mamba-Adapted SAM Framework for Medical Image Segmentation
IEEE Journal of Biomedical and Health Informatics
|March 3, 2025
Summary
This study introduces a VMamba adapter to enhance the Segment Anything Model (SAM) for medical image segmentation. The novel framework achieves state-of-the-art results, improving accuracy and efficiency in segmenting complex medical scans.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- The Segment Anything Model (SAM) demonstrates broad applicability in natural image segmentation.
- Medical image segmentation presents unique challenges due to complex anatomy and domain-specific data.
- Existing models often require specific inputs, limiting autonomous application.
Purpose of the Study:
- To develop an advanced segmentation framework for medical images by adapting the Segment Anything Model (SAM).
- To enhance the Segment Anything Model's (SAM) capability to accurately segment intricate details in medical imagery.
- To create an autonomous medical image segmentation model that minimizes manual input and improves efficiency.
Main Methods:
- Proposed a novel VMamba adapter framework integrating a lightweight Visual Mamba (VMamba) branch with the pre-trained SAM ViT encoder.
- Introduced a cross-branch attention (CBA) mechanism for effective interaction and feature fusion between SAM and VMamba.
- Developed an autonomous prediction model by removing the need for prompt-driven inputs, streamlining the segmentation workflow.
Main Results:
- The VMamba adapter framework achieved state-of-the-art performance across four medical image datasets.
- Demonstrated significant improvements on the ACDC dataset, with an average Dice coefficient increase of 0.18 and a 20.38 mm reduction in Hausdorff distance compared to AutoSAM.
- The method requires minimal additional trainable parameters, offering an efficient solution for medical image segmentation.
Conclusions:
- The proposed VMamba adapter framework effectively enhances the Segment Anything Model (SAM) for medical image segmentation tasks.
- The autonomous, efficient, and accurate segmentation capabilities represent a significant advancement for clinical applications.
- This approach offers a promising direction for leveraging large foundation models in specialized medical imaging domains.
