Abstract
Studying organ motion or pathology progression is an important task indiagnosis and therapy of various diseases. Typically, this task is approachedby deformable registration of successive images followed by the analysis of theresulting deformation field(s). Most registration methods require priorknowledge in the form of regularization of the image transformation which isoften sensitive to tuneable parameters. Alternatively, we present aregistration approach which learns a low-dimensional stochastic parametrizationof the deformation -- unsupervised, by looking at images. Hereby, spatialregularization is replaced by a constraint on this parameter space to follow aprescribed probabilistic distribution, by using a conditional variationalautoencoder (CVAE). This leads to a generative model designed to be structuredand more anatomy-invariant which makes the deformation encoding potentiallyuseful for analysis tasks like the transport of deformations. We also constrainthe deformations to be diffeomorphic using a new differentiable exponentiationlayer. We used data sets of 330 cardiac and 1000 brain images and demonstrateaccurate registration results comparable to two state-of-the-art methods.Besides, we evaluate the learned deformation encoding in two preliminaryexperiments: 1) We illustrate the model's anatomy-invariance by transportingthe encoded deformations from one subject to another. 2) We evaluate thestructure of the encoding space by clustering diseases.