Fast multiatlas selection using composition of transformations for radiation therapy planning

In radiation therapy, multiatlas segmentation is recognized as being accurate, but is generally not considered scalable since the highest accuracy is achieved only when using a large atlas database. The fundamental problem is to use such a large database, to accurately represent the population variability, while conserving a relatively small computational cost. A method based on the composition of transformations is proposed to address this issue. The main novelties and key contributions of this paper are the definition of a transitivity error function and the presentation of an image clustering scheme that is based solely on the computed registration transformations. Leave-one-out experiments conducted on a database of N = 50 MR prostate scans demonstrate that a reduction of (N - 1) = 49x in the number of pre-alignment registrations, and of 3.2x in term of total registration effort, is possible without significant impact on segmentation quality.