SCorP: Statistics-Informed Dense Correspondence Prediction Directly from Unsegmented Medical Images.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning in healthcare

Background:

• Statistical shape modeling (SSM) quantifies anatomical variability but is resource-intensive.
• Traditional SSM methods and current deep learning approaches require manual supervision and struggle with complex anatomies due to linearity assumptions.

Purpose of the Study:

• To introduce SCorP, a novel framework for unsupervised, direct prediction of surface-based correspondences from unsegmented images.
• To overcome limitations of traditional SSM and deep learning methods, including manual resource demands and linearity assumptions.

Main Methods:

• Developed SCorP, a framework leveraging unsupervised learning of shape priors from surface meshes.
• Utilized a strong shape prior to guide a student network in learning image-based features for correspondence prediction.
• Eliminated the need for optimized shape models for training supervision.

Main Results:

• SCorP accurately predicts surface-based correspondences directly from unsegmented images.
• The framework enhances the accuracy and robustness of image-driven SSM.
• Demonstrated effectiveness on LGE MRI left atrium and Abdomen CT-1K liver datasets.

Conclusions:

• SCorP offers a streamlined, unsupervised approach to SSM, reducing manual and computational burdens.
• The method alleviates the linearity assumption, enabling better modeling of complex anatomies.
• Presents a compelling, more efficient alternative to existing fully supervised SSM techniques.