Unsupervised learning of spatially varying regularization for diffeomorphic image registration

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning

Background:

• Deformable image registration is crucial for medical image analysis.
• Spatially varying regularization is vital for handling anatomical variations.
• Current deep learning models often use spatially invariant regularization, limiting performance.

Purpose of the Study:

• To develop a deep learning framework for end-to-end learning of spatially varying deformation regularizers.
• To improve the accuracy and interpretability of deformable image registration.
• To enable automatic hyperparameter tuning for registration tasks.

Main Methods:

• Proposed a hierarchical probabilistic model for learning deformation regularization strength.
• Integrated a prior distribution on regularization strength for data-driven learning.
• Utilized Bayesian optimization for automatic hyperparameter tuning.

Main Results:

• Demonstrated significant improvements in registration performance on public datasets.
• Showcased enhanced interpretability of deep learning-based registration.
• Maintained smooth deformations throughout the registration process.

Conclusions:

• The proposed method effectively learns spatially varying regularization for deformable image registration.
• It offers a flexible and integrable solution for various registration network architectures.
• The approach enhances both performance and interpretability in medical image analysis.