Learning Robust Shape Regularization for Generalizable Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Generalizable medical image segmentation is crucial for applying models to new domains.
• Existing methods often struggle with domain shifts, where texture and style variations interfere with shape extraction.
• Shape priors offer robustness but are susceptible to domain-specific interferences in deep learning models.

Purpose of the Study:

• To develop a novel framework for robust medical image segmentation across different domains.
• To address the challenge of domain-specific texture and style interferences undermining shape representations.
• To improve the generalizability and stability of segmentation models under domain shift.

Main Methods:

• A novel whitening transform-based probabilistic shape regularization extractor (WT-PSE) was developed to suppress domain-specific interferences.
• A Wasserstein distance-guided knowledge distillation scheme was employed for flexible shape extraction during inference.
• A novel instance-domain whitening transform method was incorporated for stable training and improved performance.

Main Results:

• The proposed WT-PSE effectively extracts robust and high-quality shape representations by suppressing undesirable interferences.
• The framework demonstrated improved performance in both multi-domain and single-domain generalization tasks.
• The instance-domain whitening transform facilitated a more stable training process.

Conclusions:

• Separating shape regularization from segmentation maps is a promising approach for generalizable medical image segmentation.
• The proposed WT-PSE and associated methods effectively enhance robustness and performance under domain shifts.
• This work contributes to more reliable and adaptable medical image analysis tools.