Uncertainty-based cardiac image registration using variational autoencoder with nonuniformly spaced control points.

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Variational Bayesian (VB) image registration models offer uncertainty quantification for applications like cardiac motion estimation.
• Existing VB models face challenges in extracting boundary features and effectively utilizing uncertainty information.
• Prior distributions in VB models often struggle to balance accuracy and the posterior-prior gap.

Purpose of the Study:

• To enhance Variational Bayesian image registration by addressing limitations in feature extraction and uncertainty utilization.
• To improve the accuracy of cardiac motion estimation through advanced VB image registration techniques.
• To develop a more robust VB model that effectively handles tissue boundaries and regions of interest.

Main Methods:

• Incorporation of non-uniformly spaced control points to target boundary displacements.
• Development of a network for concurrent extraction of image and spatial features from control points.
• Integration of Displacement Vector Field (DVF) uncertainty to prioritize Regions of Interest (ROIs) and enhance generative likelihood.
• Employment of a factorized prior distribution to regularize the posterior and reduce KL divergence.

Main Results:

• The proposed network outperforms state-of-the-art registration networks on four public datasets.
• The model successfully extracts features from object boundaries using non-uniformly spaced control points.
• Uncertainty-guided generative likelihood accurately matches ROIs across images.
• The factorized prior significantly improves reconstruction accuracy.

Conclusions:

• The enhanced VB image registration model effectively utilizes non-uniformly spaced control points for boundary feature extraction.
• Uncertainty-based generative likelihood accurately guides the DVF for precise ROI matching.
• The factorized prior demonstrably enhances reconstruction accuracy compared to existing methods.