LMS-Net: A learned Mumford-Shah network for binary few-shot medical image segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning

Background:

• Few-shot semantic segmentation (FSS) is crucial for data-scarce medical imaging.
• Existing FSS methods often lack interpretability and fail to capture physical structures.
• There is a need for interpretable FSS models that incorporate domain-specific knowledge.

Purpose of the Study:

• To propose a novel deep unfolding network, LMS-Net, for interpretable few-shot semantic segmentation in medical imaging.
• To integrate pixel-to-prototype comparison and deep priors within a unified framework.
• To enhance the modeling of complex spatial structures and physical properties of semantic regions.

Main Methods:

• Developed the Learned Mumford-Shah Network (LMS-Net) based on a learned Mumford-Shah model.
• Formulated the LMS model into prototype and mask update tasks, solved via alternating optimization.
• Unfolded the iterative optimization steps into network modules for interpretability.

Main Results:

• LMS-Net demonstrated superior accuracy and robustness on three medical segmentation datasets.
• The method effectively handles complex structures and challenging segmentation scenarios.
• The proposed network provides clear interpretability, a key advantage over existing FSS methods.

Conclusions:

• LMS-Net advances few-shot semantic segmentation in medical imaging by offering enhanced interpretability and performance.
• The integration of the learned Mumford-Shah model provides a strong foundation for future FSS research.
• The method shows significant potential for practical applications in medical image analysis.