Differences in morphometricity from voxel-wise and vertex-wise processing of T1w brain images

2303 

Abstract Submission 

elise delzant¹, Arya Yazdan-Panah², olivier colliot², Baptiste Couvy-Duchesne³

¹Sorbonne University, Paris Brain Institute - ICM, CNRS, Inria, Inserm, AP-HP, Paris, Paris, ²Sorbonne University, Paris Brain Institute - ICM, CNRS, Inria, Inserm, AP-HP, paris, paris, ³The University of Queensland, Brisbane, QLD

elise delzant  
Sorbonne University, Paris Brain Institute - ICM, CNRS, Inria, Inserm, AP-HP
Paris, Paris

Arya Yazdan-Panah  
Sorbonne University, Paris Brain Institute - ICM, CNRS, Inria, Inserm, AP-HP
paris, paris

olivier colliot  
Sorbonne University, Paris Brain Institute - ICM, CNRS, Inria, Inserm, AP-HP
paris, paris

Baptiste Couvy-Duchesne  
The University of Queensland
Brisbane, QLD

T1w brain MRI images provide a detailed mapping of the grey-matter structure and can be processed using different imaging software. Currently, there are no guidelines to select an image processing pipeline, and there is a growing concern that the choice of image processing could introduce a significant variability in brain representation contributing to the reproducibility crisis in neuroimaging. To progress our understanding of the effect of brain MRI processing on results, we conducted analyses utilizing five high-dimensional representations (voxel-based or vertices-based) of the grey-matter, to contrast their sensitivity to confounders and their ability to capture traits/disorders of interest.

We studied 42,272 imaged participants from the UKBiobank, divided into a discovery and replication sample, based on their assessment center. We processed all images with five different standard processing. First, FSLANAT and FSLVBM, both FSL-based processing [1] that give slightly different measurements of grey-matter density [2]. Then, SPM-based processing are implemented in the ENIGMA CAT12 toolbox [3], outputting surface-based (CAT12-Surface, measuring cortical thickness) and volume-based representations (CAT12-Volume measuring grey-matter density). Lastly, FreeSurfer [4] based processing offers surface-based representations of cortical grey-matter thickness (FreeSurfer thickness) as well as of cortical surface area, subcortical thickness and surface (FreeSurfer all modalities). We studied the association between grey-matter processing and 27 traits including 9 putative brain imaging confounders (e.g. head motion, position in scanner, time since first scan, and body size), demographics, as well as 18 traits of interest (e.g. cognition, education level, psychiatric domains or disease status). We used linear mixed models, implemented in an efficient C++ software that can deal with large dataset [6], to estimate the percentage of trait variance captured by all vertices/voxels measurements, coined "morphometricity" [5].

We found that all the putative imaging confounders exhibited a large morphometricity (20-80% of variance accounted for). However, we noted differences between processing as CAT12-Surface was less impacted by confounders (morphometricity range 18%-38%), while FreeSurfer all modalities exhibited morphometricity between 38% and 62%. As for FSLANAT and FSLVBM, they exhibited the largest associations (range 48%-76%). For example, head motion (during rs-fMRI) exhibited a morphometricity of 20% with CAT12-Surface versus 53% for FSLANAT and FSLVBM. After controlling for all confounders and demographics, we detected significant morphometricity for most traits and processing but estimates varied between traits and processing. Overall, CAT12-Surface yielded the smallest associations with traits of interests (morphometricity in 0.5%-13%). In contrast, FreeSurfer all modalities (resp. FSLANAT and FSLVBM) yielded larger associations with traits of interest (in average 2.8 -resp.3.2- times larger than CAT12-Surface, and morphometricities estimates in [3%,35%]). We confirmed the robustness of our morphometricity estimates in the replication sample.

Our results reveal that the different grey-matter processing show different levels of contamination by imaging and body size confounders, which may guide developers of processing software to investigate and reduce the source of contamination. In addition, our results highlight that these confounders should be systematically included in grey-matter analyses of the UKB. Lastly, our results provide quantified guidance on which grey-matter processing capture the most information among traits of interests (e.g. cognition, education and clinical status). Thus, FSL-based and FreeSurfer (all measurements) appear to capture more information than the other processing, enabling to construct better predictors and more complete maps of trait-associated grey-matter regions.

Genetic Modeling and Analysis Methods

Image Registration and Computational Anatomy ²

Methods Development

Cortical Anatomy and Brain Mapping

Anatomical MRI ¹

Data analysis

Degenerative Disease

Informatics

Morphometrics

Statistical Methods

STRUCTURAL MRI