Abstract
In human neuroimaging studies, atlas registration enables mapping MRI scansto a common coordinate frame, which is necessary to aggregate data frommultiple subjects. Machine learning registration methods have achievedexcellent speed and accuracy but lack interpretability. More recently,keypoint-based methods have been proposed to tackle this issue, but theiraccuracy is still subpar, particularly when fitting nonlinear transforms. Herewe propose Registration by Regression (RbR), a novel atlas registrationframework that is highly robust and flexible, conceptually simple, and can betrained with cheaply obtained data. RbR predicts the (x,y,z) atlas coordinatesfor every voxel of the input scan (i.e., every voxel is a keypoint), and thenuses closed-form expressions to quickly fit transforms using a wide array ofpossible deformation models, including affine and nonlinear (e.g., Bspline,Demons, invertible diffeomorphic models, etc.). Robustness is provided by thelarge number of voxels informing the registration and can be further increasedby robust estimators like RANSAC. Experiments on independent public datasetsshow that RbR yields more accurate registration than competing keypointapproaches, while providing full control of the deformation model.