The impact of lesion in-painting and registration methods on voxel-based morphometry in detecting regional cerebral gray matter atrophy in multiple sclerosis

Abstract

Background and purpose: VBM has been widely used to study GM atrophy in MS. MS lesions lead to segmentation and registration errors that may affect the reliability of VBM results. Improved segmentation and registration have been demonstrated by WM LI before segmentation. DARTEL appears to improve registration versus the USM. Our aim was to compare the performance of VBM-DARTEL versus VBM-USM and the effect of LI in the regional analysis of GM atrophy in MS.

Materials and methods: 3T T1 MR imaging scans were acquired from 26 patients with RRMS and 28 age-matched NC. LI replaced WM lesions with normal-appearing WM intensities before image segmentation. VBM analysis was performed in SPM8 by using DARTEL and USM with and without LI, allowing the comparison of 4 VBM methods (DARTEL + LI, DARTEL - LI, USM + LI, and USM - LI). Accuracy of VBM was assessed by using NMI, CC, and a simulation analysis.

Results: Overall, DARTEL + LI yielded the most accurate GM maps among the 4 methods (highest NMI and CC, P < .001). DARTEL + LI showed significant GM loss in the bilateral thalami and caudate nuclei in patients with RRMS versus NC. The other 3 methods overestimated the number of regions of GM loss in RRMS versus NC. LI improved the accuracy of both VBM methods. Simulated data suggested the accuracy of the results provided from patient MR imaging analysis.

Conclusions: We introduce a pipeline that shows promise in limiting segmentation and registration errors in VBM analysis in MS.

Fig 1.

Regions of decreased GM volume in patients with RRMS (n = 26) compared with NC (n = 28) (P < .05, corrected for multiple comparisons by using false discovery rate, cluster size >20), overlaid on a GM template (A–D). Comparisons of all 4 methods (DARTEL + LI, DARTEL − LI, USM + LI, and USM − LI) are illustrated. The pattern of GM atrophy in patients with RRMS shown by DARTEL is more focal, while with USM, it is more widespread. DARTEL + LI (C) likely yielded the most accurate results compared with DARTEL − LI (A), USM − LI (B), and USM + LI (D) (see “Discussion”). Bar is color-coded for t values. Images are presented in the neurologic convention (right side of image = right side of brain). See “Materials and Methods” and “Results” sections for more details.

Fig 2.

Regions of GM atrophy by using simulated ground truth data (A–D). Comparisons based on all methods are illustrated. Sixteen normal brain T1-weighted images were used as a control group and were compared with 2 groups of simulated patients (P < .05, corrected for multiple comparisons by using false discovery rate, cluster size >20). The results are displayed on a 3D glass brain. In 1 group (C and D), we only simulated atrophy of the thalamus and caudate, and in the other (A and B), we simulated atrophy in the same regions and added artificial lesions. DARTEL in the presence of only atrophy (C) yielded the most accurate results, detecting the atrophy under the ground truth and reducing the number of FPs compared with USM (D). The presence of lesions increased the number of FP errors in both DARTEL and USM (A and B). See “Materials and Methods” and “Results” section for more details.