Improving spatial normalization of brain diffusion MRI to measure longitudinal changes of tissue microstructure in the cortex and white matter
ABSTRACTBackgroundFractional anisotropy (FA) and mean diffusivity (MD) are frequently used to evaluate longitudinal changes in white matter microstructure. Recently, there has been a growing interest in identifying experience-dependent plasticity in gray matter using MD. Improving registration has thus become a major goal to enhance the detection of subtle longitudinal changes in cortical microstructure.PurposeTo optimize normalization to improve registration in gray matter and reduce variability associated with multi-session registrations.Study TypeProspective longitudinal studySubjectsTwenty-one healthy subjects (18-31 years old) underwent 9 magnetic resonance imaging (MRI) scanning sessions each.Field Strength/Sequence3.0T, diffusion-weighted multiband-accelerated sequence, MP2RAGE sequence.AssessmentDiffusion-weighted images were registered to standard space using different pipelines that varied in the features used for normalization, namely the non-linear registration algorithm (FSL vs ANTs), the registration target (FA-based vs T1-based templates), and the use of intermediate individual (FA-based or T1-based) targets. We compared the across-session test-retest reproducibility error from these normalization approaches for FA and MD in white and gray matters.Statistical TestsReproducibility errors were compared using a repeated-measures analysis of variance with pipeline as within-subject factor.ResultsThe registration of FA data to the FMRIB58 FA atlas using ANTs yielded lower reproducibility errors in white matter (p<0.0001) with respect to FSL. Moreover, using the MNI152 T1 template as the target of registration resulted in lower reproducibility errors for MD (p<0.0001), whereas the FMRIB58 FA template performed better for FA (p<0.0001). Finally, the use of an intermediate individual template improved reproducibility when registration of the FA images to the MNI152-T1 was carried out within modality (FA-FA) (p<0.05), but not via a T1-based individual template.Data ConclusionA normalization approach using ANTs to register FA images to the MNI152 T1 template via an individual FA template minimized test-retest reproducibility errors both for gray and white matter.