Learnable Prompting SAM-Induced Knowledge Distillation for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Limited labeled data hinders medical image segmentation accuracy.
• General segmentation models like Segment Anything Model (SAM) show promise but degrade in medical applications.
• Semi-supervised learning is crucial for overcoming data scarcity in medical image segmentation.

Purpose of the Study:

• To develop an effective semi-supervised medical image segmentation framework using knowledge distillation from SAM.
• To improve the performance of general segmentation models on specialized medical imaging tasks.
• To address the challenge of performance degradation when applying large-scale models to medical data.

Main Methods:

• Proposed a learnable prompting SAM-induced Knowledge distillation framework (KnowSAM).
• Implemented a Multi-view Co-training (MC) strategy with two distinct sub-networks.
• Introduced a Learnable Prompt Strategy (LPS) with an adapter for SAM fine-tuning and SAM-induced Knowledge Distillation (SKD).

Main Results:

• KnowSAM demonstrated superior performance compared to state-of-the-art semi-supervised segmentation approaches across various medical tasks.
• The proposed SAM distillation framework effectively transfers knowledge from SAM to sub-networks, mitigating pseudo-label errors.
• Inter-module information exchange was facilitated through SAM mask prompts generated by sub-network predictions.

Conclusions:

• KnowSAM offers a robust and effective solution for semi-supervised medical image segmentation.
• The SAM distillation framework can be integrated into existing semi-supervised methods to boost performance.
• The study highlights the potential of leveraging large-scale general models for specialized medical imaging tasks.