Not All Regions are Equal in Registration: A Segmentation-Assisted Registration Network

Existing learning-based image registration methods learn a network that outputs a spatial transformation to warp a source image to align to the target image, which is carried out through optimizing matching similarities. Such matching criteria are mostly agnostic of regions. However, in many applications of medical image registration, the correspondence with specific regions of interest (ROIs) such as organs in source and target images is more important than the whole images as not all regions are equal. Motivated by this, we hereby propose a segmentation-assisted registration network to emphasize the ROIs more. We promote the estimation of registration fields by encoding the anatomical structure extracted by manual or predicted segmentation label generated from fusing the outputs of a multi-atlas-based registration model and a deep segmentation network which is trained using a few manual labeled images. The experiments on 3D brain MRI indicate that our approach improves the registration performance with few manual segmentation labels or ROIs. Our approach also outputs competitive segmentation results.