Inter-Phase 4D Cardiac MRI Registration With a Motion Prior Derived From CTA

Area of Science:

• Medical imaging
• Cardiovascular imaging
• Image registration

Background:

• Accurate 4D cardiac MRI registration is crucial for image reconstruction and dynamic analysis.
• Complex motion and limited image quality pose challenges for traditional regularization methods.

Purpose of the Study:

• To introduce a motion representation model (MRM) into a registration network for cardiac MRI.
• To impose customized, site-specific, and spatially variant priors for cardiac motion.
• To improve the accuracy and efficiency of 4D cardiac MRI registration.

Main Methods:

• A novel approach regularizes deep registration using a deformation vector field (DVF) representation model derived from computed tomography angiography (CTA).
• The MRM, implemented as a convolutional auto-encoder, captures spatially variant DVF Jacobian patterns.
• The CTA-derived MRM is integrated into an unsupervised network for MRI registration, tested on various cardiac MRI scans.

Main Results:

• The MRM significantly improved registration accuracy, with Hausdorff distances of 2.23 mm (left ventricle), 7.21 mm (right ventricle), and 4.42 mm (pulmonary artery).
• Landmark registration error was 2.23 mm, comparable to SimpleElastix but with a substantial reduction in registration time (0.02 s vs. 40 s).
• The MRM demonstrated robustness across different DVF sample generation methods.

Conclusions:

• The developed model achieves high accuracy and efficiency, surpassing the quality limitations of standard MRI.
• Its robustness to training DVF generation schemes makes it adaptable to diverse clinical data and software resources.