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

The Variational Bayesian (VB) image registration model has garnered recent attention for its ability to offer uncertainty, particularly in the context of cardiac motion estimation. Nonetheless, several challenges have plagued VB image registration. Firstly, the Convolutional Neural Networks (CNNs) in VB excel with grid-based image features make it challenging to extract features from non-uniformly located points located at tissue boundaries. Secondly, the underutilization of VB-provided uncertainty and the misfocus of the regions of interest (ROIs) lead to misleading generative likelihoods. Lastly, existing VB prior distributions struggle to balance the posterior-prior gap and reconstruction accuracy.

To address these concerns, we extend the VB image registration model by incorporating non-uniformly spaced control points that specifically target displacements at object boundaries. We develop a network capable of concurrently extracting image and spatial features from non-uniformly spaced control points. The uncertainty of the Displacement Vector Field (DVF) is integrated to prioritize matching ROIs in two images and enhance generative likelihood. Additionally, Our VB model employs a factorized prior to regularize the posterior distribution. Theoretical analysis shows that the factorized prior reduces the KL divergence between the posterior and the prior.

Results on four public datasets demonstrate that our network outperforms state-of-the-art registration networks while providing valuable uncertainty information on registration outcomes.

Experiments confirm that our VB image registration utilizing non-uniformly spaced control points effectively extracts features from boundary, and the uncertainty-based generative likelihood guides the DVF to match ROIs accurately across images. The factorized prior significantly improves reconstruction accuracy compared to existing priors.