Related Experiment Video
Updated: Jun 9, 2025
06:48
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
8.8K
Point-supervised Brain Tumor Segmentation with Box-prompted Medical Segment Anything Model
1Yale University, Radiology and Biomedical Imaging, New Haven, Connecticut, United States of America.
IEEE Nuclear Science Symposium Conference Record. Nuclear Science Symposium
|October 24, 2024
Summary
This study introduces a novel iterative framework for point-supervised medical image segmentation (PSS) using MedSAM. The method enhances segmentation accuracy by converting point annotations into semantic bounding boxes, improving upon traditional PSS techniques.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Accurate delineation of anatomical structures and lesions is crucial for image-guided interventions.
- Point-supervised medical image segmentation (PSS) offers a promising solution to reduce the burden of expert labeling.
- Current PSS methods struggle with precise boundary and size guidance, limiting their effectiveness.
Purpose of the Study:
- To develop an effective point-supervised medical image segmentation framework leveraging foundational vision models.
- To address the limitations of point annotations in semantic ambiguity and boundary definition.
- To improve the performance of MedSAM for point-prompted segmentation tasks.
Main Methods:
- Introduced an iterative framework for semantic-aware point-supervised MedSAM.
- Developed a semantic box-prompt generator (SBPG) to convert point inputs into refined pseudo bounding box suggestions using prototype-based semantic similarity.
- Employed a prompt-guided spatial refinement (PGSR) module to infer segmentation masks and iteratively update box proposals, harnessing MedSAM's generalizability.
Main Results:
- The proposed framework demonstrated progressively improved performance with adequate iterations.
- Evaluated on BraTS2018 for whole brain tumor segmentation.
- Achieved superior performance compared to traditional PSS methods and performance on par with box-supervised methods.
Conclusions:
- The iterative framework effectively facilitates semantic-aware point-supervised MedSAM.
- The SBPG and PGSR modules enhance the utilization of point annotations for medical image segmentation.
- The approach shows significant potential for improving the efficiency and accuracy of medical image segmentation in clinical applications.
