Mixture-of-Shape-Experts (MoSE): End-to-End Shape Dictionary Framework to Prompt SAM for Generalizable Medical...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Single domain generalization (SDG) is crucial for medical image segmentation across diverse data.
• Existing dictionary learning methods struggle with representational power and compatibility with foundation models like SAM.
• Robust shape priors are essential for improving generalization in medical image segmentation.

Purpose of the Study:

• To propose a novel Mixture-of-Shape-Experts (MoSE) framework for efficient and robust shape prior encoding in SDG.
• To integrate MoSE with the Segment Anything Model (SAM) for enhanced generalization capabilities.
• To develop an end-to-end trainable framework for medical image segmentation.

Main Methods:

• Developed a Mixture-of-Shape-Experts (MoSE) framework integrating mixture-of-experts (MoE) training with dictionary learning.
• Conceptualized dictionary atoms as 'shape experts' for encoding distinct semantic shape information.
• Utilized a gating network with SAM-guided sparse activation to fuse shape experts and prevent overfitting, providing a shape map as a prompt to SAM.

Main Results:

• The MoSE framework demonstrated effective capture of diverse and robust shape priors.
• Bidirectional integration with SAM leveraged its generalization capabilities.
• Extensive experiments on multiple public datasets confirmed the method's effectiveness.

Conclusions:

• The proposed MoSE framework offers an efficient and robust solution for leveraging semantic shape priors in medical image segmentation.
• MoSE enhances generalization by integrating with foundation models like SAM.
• The end-to-end trainable framework shows significant promise for improving medical image segmentation tasks.