Classical deformable registration techniques achieve impressive results and offer a rigorous theoretical treatment, but are computationally intensive since they solve an optimization problem for each image pair. Recently, learning-based methods have facilitated fast registration by learning spatial deformation functions. However, these approaches use restricted deformation models, require supervised labels, or do not guarantee a diffeomorphic (topology-preserving) registration. Furthermore, learning-based registration tools have not been derived from a probabilistic framework that can offer uncertainty estimates. In this paper, we build a connection between classical and learning-based methods. We present a probabilistic generative model and derive an unsupervised learning-based inference algorithm that uses insights from classical registration methods and makes use of recent developments in convolutional neural networks (CNNs). We demonstrate our method on a 3D brain registration task for both images and anatomical surfaces, and provide extensive empirical analyses. Our principled approach results in state of the art accuracy and very fast runtimes, while providing diffeomorphic guarantees. Our implementation is available at http://voxelmorph.csail.mit.edu.Keywords medical image registration · diffeomorphic registration · invertible registration · probabilistic modeling · convolutional neural networks · variational inference · machine learning arXiv:1903.03545v2 [cs.CV] 23 Jul 2019 experiments, we add baselines, new experiments on registration of both images and surfaces, and provide an analysis of the effect of our diffeomorphic implementation on field regularity and runtime. We implement our method as part of the registration framework called VoxelMorph, which is available at http://voxelmorph.csail.mit.edu.
Related Works
Classical Registration MethodsClassical methods solve an optimization over the space of deformations [5,7,8,11,19,28,66,69,70]. Common representations are displacement vector fields, including elastic-type models [8,21,62], free-form deformations with b-splines [61], statistical parametric mapping [6], Demons [56,66], and more recently discrete methods [19,30,28].Constraining the allowable transformations to diffeomorphisms ensures certain desirable properties, such as preservation of topology. Diffeomorphic transforms have seen extensive methodological development, yielding state-ofthe-art tools, such as Large Diffeomorphic Distance Metric Mapping (LDDMM) [11,14,15,32,37,49,55,70], DARTEL [5], diffeomorphic Demons [67], and symmetric normalization (SyN) [7]. In general, these tools demand substantial time and computational resources for a given image pair.Some recent GPU-based iterative algorithms use these frameworks to develop faster algorithms by requiring a GPU to be available for each registration [51,50]. Recent learning-based registration methods have demonstrated that they can provide good initializations to iterative GPU methods [10] to further improve runtime.