Registration by Regression (RbR): a framework for interpretable and flexible atlas registration.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Machine Learning

Background:

• Atlas registration is crucial for aggregating human neuroimaging data.
• Current machine learning methods lack flexibility and interpretability.
• Keypoint-based methods show promise but have suboptimal accuracy for nonlinear transforms.

Purpose of the Study:

• Introduce Registration by Regression (RbR), a novel, robust, and flexible atlas registration framework.
• Enable accurate mapping of MRI scans to a common coordinate frame.
• Provide a method adaptable to various deformation models and usable for pretraining.

Main Methods:

• RbR predicts atlas coordinates for every voxel, treating each as a keypoint.
• Utilizes closed-form expressions for rapid fitting of diverse deformation models (affine, B-spline, Demons, diffeomorphic).
• Incorporates robust estimators like RANSAC to enhance registration accuracy.

Main Results:

• RbR demonstrates high robustness and flexibility in atlas registration.
• Achieves superior accuracy compared to existing keypoint-based methods across multiple deformation models.
• Validated on independent public neuroimaging datasets.

Conclusions:

• RbR presents a significant advancement in atlas registration for neuroimaging.
• Offers a versatile and accurate solution for MRI data aggregation.
• Potential for pretraining and broader applications in medical image analysis.