A segment anything model-guided and match-based semi-supervised segmentation framework for medical imaging.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised segmentation utilizes limited labeled data alongside unlabeled data for improved model training.
• Match-based frameworks rely on consistency constraints but struggle with low-quality pseudo-labels.
• The Segment Anything Model (SAM) offers strong generalization due to extensive pre-training.

Purpose of the Study:

• To integrate the Segment Anything Model (SAM) into Match-based frameworks to enhance semi-supervised medical image segmentation.
• To improve the quality of pseudo-labels generated in semi-supervised learning using SAM's capabilities.
• To leverage SAM's generalization power for better performance across diverse medical imaging domains.

Main Methods:

• Propose SAM-Match, a novel framework combining SAM and Match-based approaches for semi-supervised medical image segmentation.
• Utilize pre-trained Match-based models to generate high-confidence predictions for prompt creation.
• Employ a fine-tuned SAM-based method with generated prompts and unlabeled data to produce high-quality pseudo-labels for training.

Main Results:

• SAM-Match achieved robust segmentation performance on cardiac MRI (ACDC), breast ultrasound (BUSI), and liver MRI datasets.
• Demonstrated high Dice scores with minimal labeled data: 89.36% on ACDC (3 labels), 59.35% on BUSI (30 labels), and 80.04% on MRLiver (3 labels).
• Statistical significance confirmed improvements over existing methods via Wilcoxon signed-rank tests.

Conclusions:

• The SAM-Match framework effectively addresses challenges in prompt generation for SAM and pseudo-labeling for Match-based models.
• Shows significant potential for accelerating semi-supervised learning adoption in medical imaging, especially in data-scarce situations.
• Code and data will be publicly released to facilitate further research and application.