Learning a Probabilistic Model for Diffeomorphic Registration
J. Krebs, H. Delingette, B. Mailhé, N. Ayache, T. Mansi
IEEE Trans Med Imaging., vol. 38, issue 2, February 2019
Deformable image registration is an essential task in medical image analysis. It describes the process of finding correspondences in a pair of images, the moving and fixed image. This mapping – the deformation field – can be used for example to find the same structures in images from different modalities or to evaluate the progression of a disease. In the presence of moving organs, registration helps in the analysis of geometric changes in successive images such as those acquired during the cardiac cycle.
We propose to learn in an unsupervised manner a low-dimensional deformation model from pairs of images which can be used for diffeomorphic registration and the analysis of deformations. An encoder-decoder framework maps similar deformations close to each other in a probabilistic encoding latent space. These z-codes enable to compare deformations and to cluster them according to various pathologies. Our framework based on variational inference also allows to generate realistic deformations and to transport them from one patient to another.
Besides, we show that the z-codes can be used to cluster cardiac diseases (classification accuracy of 83%) and to simulate a pathological cardiac motion in a healthy image.