SAM-guided Cross Pseudo Supervision Learning for Semi-supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Accurate lesion identification in medical images is vital for diagnosis and treatment.
• Limited annotations and complex lesion features hinder current segmentation model performance.
• The Segment Anything Model (SAM) shows promise in natural image segmentation but struggles with medical data.

Purpose of the Study:

• To develop a SAM-assisted semi-supervised framework for accurate lesion segmentation, especially with scarce annotations.
• To overcome the performance limitations of SAM in medical imaging by incorporating domain-specific strategies.
• To improve the representational capacity of models in the face of limited annotated medical data.

Main Methods:

• A dual-branch architecture utilizing fine-tuned SAM (ViT-Base) models with adapters.
• Development of a point prompts generator using pseudo-labels from dual-branch outputs for automatic prompt creation.
• Leveraging a pre-trained SAM (ViT-Huge) model with generated prompts to refine segmentation masks for additional supervision.

Main Results:

• The proposed framework achieves promising performance in lesion segmentation.
• Substantial improvements demonstrated over existing baseline methods in experiments.
• The method shows effectiveness even in highly annotation-scarce clinical scenarios.

Conclusions:

• The novel SAM-assisted framework effectively addresses the challenge of limited annotations in medical lesion segmentation.
• The autonomous generation of point prompts for unlabeled data enhances segmentation accuracy.
• This approach offers a viable solution for clinical practice where data annotation is resource-intensive.