Room: Track 2
Purpose: Image registration which estimates deformable vector field (DVF) from multiple image inputs is important in medical physics. In conventional registration framework, the complexity of the DVF parametrization function (such as B-spline order) and universal prescription of regularization make it challenging to accommodate heterogeneous scale levels in DVF and intensity characteristics. It also impedes translation of registration approach in different sites or domains. We hypothesize learning a DVF representation would bridge this gap and propose a novel framework to incorporate a learned flexible parametrization model that imposes implicit feasibility constraints on deformation.
Methods: A generator network is trained in an unsupervised setting to maximize the likelihood of observing moving and fixed image pairs, using an alternating back-propagation approach. The trained generative model serves as an effective low dimensional parametrization and imposes constraints on deformation. During registration, optimization is performed over this learned parametrization and does not require explicit regularization. The proposed method was tested on simulated chest CT images with dense ground-truth DVFs and landmarks to quantify performance and on 3D cardiac MRIs with left ventricle segmentations for evaluation.
Results: Experiments with simulated CT images demonstrated the efficacy of our method. Target registration error was on millimeter level, with statistically significant (t-test) improvement over all Elastic options, DIRnet and Voxelmorph. Experiments with cardiac MRIs showed that the method encouraged physical and physiological feasibility of deformation and achieved a dice of (0.93+/-0.03) with statistically significant (Wilcoxon) improvement over all Elastic options and both variations of DIRnet. Mean average surface distance was on millimeter level (subpixel), comparable to the best Elastix setting without statistical significance. The average 3D registration time was 12.78s, faster than 24.70s in Elastix.
Conclusion: The learned implicit parametrization could be an efficacious alternative to regularized B-spline model, that is more flexible in admitting spatial heterogeneity.