Learning-based Neuroimage Registration

Abstract

Neuroimage registration has been a crucial area of research in medical image analysis for many years. Aligning brain images of different subjects in such a way that same anatomical structures correspond spatially is required in many different applications, including neuroimage classification, computer aided diagnosis, statistical quanti?ation of human brains and neuroimage segmentation. We combine statistical learning, computer vision and medical image analysis to propose a multiresolution framework for learning-based neuroimage registration. Our approach has four distinct characteristics not present in other registration methods. First, instead of subjectively choosing which features to use for registration, we employ feature selection at di?rent image scales to learn an appropriate subset of features for registering a specific pair of neuroimages. Second, we use interesting-voxel selection to identify image voxels that have the most distinct image feature vectors. These voxels are then used to estimate the deformation field for registration. Third, we iteratively improve our choice of features and interesting voxels during registration process. Fourth, we create and take advantage of a statistical model containing information on image feature distributions in each anatomical location.