Bayesian Unsupervised Disentanglement of Anatomy and Geometry for Deep Groupwise Image Registration.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning

Background:

• Groupwise image registration aligns multiple images simultaneously.
• Existing methods often rely on complex similarity measures and lack interpretability.
• Multi-modal registration presents challenges due to differing image contrasts.

Purpose of the Study:

• To develop a general Bayesian learning framework for unsupervised multi-modal groupwise image registration.
• To disentangle common anatomy and geometric variations using latent variables.
• To achieve accurate and interpretable registration through hierarchical Bayesian inference.

Main Methods:

• A novel hierarchical variational auto-encoding architecture for inference.
• Probabilistic modeling of the image generative process.
• Unsupervised closed-loop self-reconstruction for learning registration parameters.
• Disentanglement learning for capturing latent anatomical structures.

Main Results:

• The proposed framework achieves superior accuracy, efficiency, scalability, and interpretability compared to conventional methods.
• Demonstrated effectiveness across diverse medical imaging datasets (cardiac, brain, abdominal).
• Inferred structural representations capture visual semantics from multi-modal data.

Conclusions:

• The Bayesian learning framework offers a robust and unsupervised approach to multi-modal groupwise image registration.
• The hierarchical variational auto-encoding method provides interpretable registration parameters.
• The framework is scalable for large-scale image groups and adaptable to variable sizes.