A Deep Learning Network for Coarse-to-Fine Deformable Medical Image Registration

Deformable image registration was a fundamental task in medical image analysis. Recently published registration methods based on deep learning have shown promising results. However, these algorithms brought limited precision improvement due to the similar learning framework. In order to address this shortcoming, we proposed a novel two-stage framework for deep learning based image registration. The new network computed deformation fields on different scales, similar to methods using auto-context strategy. Thereby, a coarse-scale alignment was obtained by the first half part of our network, which was subsequently improved on finer scale by the second half part. The new model could directly estimate the final deformation field in an end to end way. We demonstrated our method on the task of brain magnetic resonance (MR) image registration and showed that the new model could reach significantly better registration results.