Point-supervised Brain Tumor Segmentation with Box-prompted MedSAM.

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, but often lacks precision.
• Existing foundational models like MedSAM excel with bounding-box prompts but struggle with point annotations and semantic ambiguity.

Purpose of the Study:

• To develop an iterative framework for semantic-aware point-supervised segmentation using MedSAM.
• To address the limitations of point annotations in medical image segmentation.
• To improve the effectiveness of PSS by leveraging foundational models.

Main Methods:

• Introduction of a semantic box-prompt generator (SBPG) to convert point inputs into pseudo bounding box suggestions.
• Refinement of bounding box suggestions using prototype-based semantic similarity.
• Utilization of a prompt-guided spatial refinement (PGSR) module with MedSAM for mask inference and iterative improvement of box proposals.

Main Results:

• The proposed framework demonstrates progressively improved performance with iterations.
• Evaluation on the BraTS2018 dataset for whole brain tumor segmentation showed superior results compared to traditional PSS methods.
• The method achieved performance on par with box-supervised segmentation techniques.

Conclusions:

• The developed iterative framework effectively enhances semantic-aware point-supervised segmentation with MedSAM.
• This approach offers a viable solution for precise medical image segmentation using limited point annotations.
• The findings suggest a promising direction for improving automated segmentation in medical imaging applications.