Effect of non-rigid registration algorithms on the analysis of brain MR images with deformation based morphometry

Abstract

Deformation Based Morphometry (DBM) is a widely used method for characterizing anatomical differences across populations. DBM is based on the analysis of the deformation fields generated by non-rigid registration algorithms that warp the individual volumes to a DBM atlas. Although several studies have compared non-rigid registration algorithms for segmentation tasks, few studies have compared the effect of the registration algorithms on group differences that may be uncovered through DBM. The overarching goal of this dissertation is to assess qualitatively and quantitatively the extent to which DBM results are a function of the registration algorithms used to compute the deformation fields. Five well-established non-rigid registration algorithms are compared and tested on two different real data sets and a series of simulated datasets. The first real data set has large and well-documented anatomical differences between normal subjects and subjects with the Williams Syndrome. The second real data set contains MR images of third-grade children with different levels of mathematical abilities. Anatomical differences in this data set are more subtle. Because the lack of ground truth makes it difficult to compare algorithms, a series of simulated MR images with various known anatomical differences are produced. DBM results obtained with the five registration algorithms are compared with the introduced ground truth both qualitatively and quantitatively. The main conclusions that can be drawn from this work are that (1) DBM-based findings are indeed dependent on the registration algorithm that is used and (2) the Family Wise Error (FWE) statistical scheme that is commonly used for multiple comparison correction may be over-conservative. When performing DBM analysis, our suggestions would be to use more than one algorithm and to look for regions that are consistently labeled as statistically significant across these algorithms. We are also the first to report that brain anatomy may correlate with the level of mathematical performances in a relatively large population of third graders.