Rigid Registration Method for Medical Volumes with Large Deformations and Missing Data

A rigid registration is a crucial initial step for a correct deformable medical image registration. In this work, we propose rigid registration method resistant to large deformations and missing data. The proposed method is based on the bones segmentation, feature matching and outliers elimination inspired by traditional computer vision approach. The method is compared to other state-of-the-art algorithms, the iterative closest point and intensity-based registration using widely available dataset. The proposed algorithm does not fail into local minima and reconstructs correct deformations for average vector length greater than 150 mm and data overlap ratio less than 50%, where currently applied methods fail. The algorithm is evaluated using angle and magnitude errors between corresponding deformation vectors, Hausdorff distance between bone segmentations and resistance to fail into local minima.