An automated registration algorithm for measuring MRI subcortical brain structures

Iosifescu DV, Shenton ME, Warfield SK, Kikinis R, Dengler J, Jolesz FA, McCarley RW

Neuroimage 1997 Jul;6(1):13-25

PMID: 9245652

Abstract

An automated registration algorithm was used to elastically match an anatomical magnetic resonance (MR) atlas onto individual brain MR images. Our goal was to evaluate the accuracy of this procedure for measuring the volume of MRI brain structures. We applied two successive algorithms to a series of 28 MR brain images, from 14 schizophrenia patients and 14 normal controls. First, we used an automated segmentation program to differentiate between white matter, cortical and subcortical gray matter, and cerebrospinal fluid. Next, we elastically deformed the atlas segmentation to fit the subject’s brain, by matching the white matter and subcortical gray matter surfaces. To assess the accuracy of these measurements, we compared, on all 28 images, 11 brain structures, measured with elastic matching, with the same structures traced manually on MRI scans. The similarity between the measurements (the relative difference between the manual and the automated volume) was 97% for whole white matter, 92% for whole gray matter, and on average 89% for subcortical structures. The relative spatial overlap between the manual and the automated volumes was 97% for whole white matter, 92% for whole gray matter, and on average 75% for subcortical structures. For all pairs of structures rendered with the automated and the manual method, Pearson correlations were between r = 0.78 and r = 0.98 (P < 0.01, N = 28), except for globus pallidus, where r = 0.55 (left) and r = 0. 44 (right) (P < 0.01, N = 28). In the schizophrenia group, compared to the controls, we found a 16.7% increase in MRI volume for the basal ganglia (i.e., caudate nucleus, putamen, and globus pallidus), but no difference in total gray/white matter volume or in thalamic MR volume. This finding reproduces previously reported results, obtained in the same patient population with manually drawn structures, and suggests the utility/efficacy of our automated registration algorithm over more labor-intensive manual tracings.