Related Experiment Video
Updated: May 23, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
WeakMedSAM: Weakly-Supervised Medical Image Segmentation via SAM With Sub-Class Exploration and Prompt Affinity
IEEE Transactions on Medical Imaging
|March 10, 2025
Summary
This study introduces WeakMedSAM, a novel weakly-supervised model that reduces medical image annotation costs by leveraging the Segmenting Anything Model (SAM). It achieves accurate segmentation with less data, improving efficiency in medical AI.
Area of Science:
- Computer Vision
- Medical Image Analysis
- Artificial Intelligence
Background:
- Foundation models show significant progress in vision tasks.
- Existing methods often require extensive pixel-wise annotations for medical image segmentation using models like the Segmenting Anything Model (SAM).
- High labeling costs hinder the widespread application of advanced segmentation models in medicine.
Purpose of the Study:
- To develop a weakly-supervised SAM-based segmentation model (WeakMedSAM) to reduce the need for extensive medical image annotations.
- To improve the accuracy and efficiency of medical image segmentation while minimizing labeling effort.
- To investigate novel modules for feature representation and segmentation refinement within a SAM-like framework.
Main Methods:
- Proposed WeakMedSAM incorporates a sub-class exploration module to learn accurate feature representations and mitigate co-occurrence issues in medical images.
- A prompt affinity mining module is introduced to enhance class activation maps by utilizing SAM's prompt capabilities for random-walk refinement.
- The method is designed to be compatible with various SAM-like backbones, demonstrated with SAMUS and EfficientSAM.
Main Results:
- WeakMedSAM demonstrated promising segmentation performance on three benchmark datasets: BraTS 2019, AbdomenCT-1K, and MSD Cardiac.
- The model effectively reduces the reliance on fully supervised, pixel-wise annotated data.
- Experimental results validate the efficacy of the proposed sub-class exploration and prompt affinity mining modules.
Conclusions:
- WeakMedSAM offers an effective weakly-supervised approach for medical image segmentation, significantly reducing annotation costs.
- The proposed modules enhance feature representation and segmentation quality, making advanced models more accessible for medical applications.
- This work paves the way for more efficient and cost-effective AI-driven medical image analysis.
