Peeking into the Black Box of Coregistration in Clinical fMRI: Which Registration Methods Are Used and How Well Do They Perform?

Abstract

BACKGROUND AND PURPOSE:Interpretation of fMRI depends on accurate functional-to-structural alignment. This study explores registration methods used by FDA-approved software for clinical fMRI and aims to answer the following question: What is the degree of misalignment when registration is not performed, and how well do current registration methods perform? MATERIALS AND METHODS:This retrospective study of presurgical fMRI for brain tumors compares nonregistered images and 5 registration cost functions: Hellinger, mutual information, normalized mutual information, correlation ratio, and local Pearson correlation. To adjudicate the accuracy of coregistration, we edge-enhanced echo-planar maps and rated them for alignment with structural anatomy. Lesion-to-activation distances were measured to evaluate the effects of different cost functions. RESULTS:Transformation parameters were congruent among Hellinger, mutual information, normalized mutual information, and the correlation ratio but divergent from the local Pearson correlation. Edge-enhanced images validated the local Pearson correlation as the most accurate. Hellinger worsened misalignment in 59% of cases, primarily exaggerating the inferior translation; no cases were worsened by the local Pearson correlation. Three hundred twenty lesion-to-activation distances from 25 patients were analyzed among nonregistered images, Hellinger, and the local Pearson correlation. ANOVA analysis revealed significant differences in the coronal (P Ͻ .001) and sagittal (P ϭ .04) planes. If registration is not performed, 8% of cases may have a Ͼ3-mm discrepancy and up to a 5.6-mm lesion-toactivation distance difference. If a poor registration method is used, 23% of cases may have a Ͼ3-mm discrepancy and up to a 6.9-mm difference. CONCLUSIONS:The local Pearson correlation is a special-purpose cost function specifically designed for T2*-T1 coregistration and should be more widely incorporated into software tools as a better method for coregistration in clinical fMRI. ABBREVIATIONS: AFNI ϭ Analysis of Functional Neuro Images; CR ϭ correlation ratio; HEL ϭ Hellinger; LAD ϭ lesion-to-activation distance; LPC ϭ local Pearson correlation; MI ϭ mutual information; NMI ϭ normalized mutual information; NR ϭ nonregistered

Improving 3D edge detection for visual inspection of MRI coregistration and alignment

“…Automated methods do not always perform as intended (Raslau et a., 2018), and reliance on automated methods for quality assurance is problematic. Indeed, a grossly distorted solution caused by the presence of a lesion (Brett et al, 2001) or susceptibility artifacts (Raslau et al, 2018) can often have better alignment with these cost functions than solutions that are actually more accurate. Therefore, having reliable edge detection and display algorithms is essential for neuroimaging interpretation and processing.…”