A comparative study of rigid-body registration algorithms for the alignment of longitudinal structural MRI of the brain.

BACKGROUND

Longitudinal structural MRI (sMRI) may be used to characterize brain morphological changes over time. A key requirement is accurate rigid-body alignment of longitudinal sMRI.

NEW METHOD: We have developed the automatic temporal registration algorithm (ATRA) for this purpose. ATRA is a landmark-based approach capable of registering dozens of sMRI simultaneously.

COMPARISON WITH EXISTING

METHODS

The aim of this research was to evaluate the accuracy and inverse-consistency of ATRA in comparison to FSL, FreeSurfer, and ANTS. Absent a ground truth, it is only possible to determine the degree of discrepancy between algorithms. We propose that if the discrepancy exceeds a certain threshold, the relative accuracy of algorithms could be determined visually. We computed the discrepancy between ATRA and each of the other three methods for the alignment of 150 pairs of longitudinal sMRI.

RESULTS

We visually rated the accuracy of alignments in cases where the discrepancy was greater than .5 mm while the rater was agnostic to the registration method. In those instances, ATRA was considered more accurate than FSL in 46 out of 48 cases (p < .0001), more accurate than FreeSurfer in 6 out of 7 cases (p = .0625), and more accurate than ANTS in all 6 cases (p < .05). ATRA was also significantly more inverse-consistent than all other methods. FreeSurfer was more inverse-consistent than FSL and ANTS, while ANTS outperformed FSL (p < .0001 for all 6 pairwise comparisons).

CONCLUSIONS

In addition to being capable of performing unbiased group-wise registration, ATRA is the most accurate algorithm in comparison to several commonly used rigid-body alignment methods.