SAM-driven cross prompting with adaptive sampling consistency for semi-supervised medical image segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Semi-supervised learning (SSL) is crucial for medical image segmentation due to limited labeled data.
• Visual foundation models like Segment Anything Model (SAM) offer improved sample efficiency.
• Harnessing foundation models for SSL in medical imaging remains a challenge.

Purpose of the Study:

• To propose a novel SAM-driven framework (CPAC-SAM) for semi-supervised medical image segmentation.
• To enhance learning from both labeled and unlabeled data using cross prompting and adaptive sampling.
• To improve prompt reliability and consistency for robust segmentation.

Main Methods:

• Developed a SAM-driven cross prompting framework with a dual-branch structure.
• Implemented a prototype-guided grid sampling strategy for adaptive prompt generation.
• Introduced prompt consistency regularization to reduce SAM's sensitivity.

Main Results:

• CPAC-SAM demonstrated superior performance over state-of-the-art SSL methods across five medical image segmentation tasks (2D and 3D).
• Achieved significant Dice improvements, including over 4.1% for breast cancer and 3.8% for left atrium segmentation.
• Validated effectiveness across various labeled-data ratios and modalities.

Conclusions:

• CPAC-SAM effectively integrates foundation models into semi-supervised medical image segmentation.
• The proposed cross prompting, adaptive sampling, and consistency regularization enhance learning efficiency and accuracy.
• This framework offers a promising approach for advancing medical image analysis with limited labeled data.