Choice of Voxel-based Morphometry processing pipeline drives variability in the location of neuroanatomical brain markers

ARTICLE https://doi.org/10.1038/s42003-022-03880-1 OPEN Choice of Voxel-based Morphometry processing pipeline drives variability in the location of neuroanatomical brain markers 1234567890():,; Xinqi Zhou 1,2,4 ✉, Renjing Wu1, Yixu Zeng1, Ziyu Qi1, Stefania Ferraro1,3, Lei Xu1,2, Xiaoxiao Zheng1, Jialin Li1, Meina Fu1, Shuxia Yao 1, Keith M. Kendrick1 & Benjamin Becker 1,4 ✉ Fundamental and clinical neuroscience has benefited tremendously from the development of automated computational analyses. In excess of 600 human neuroimaging papers using Voxel-based Morphometry (VBM) are now published every year and a number of different automated processing pipelines are used, although it remains to be systematically assessed whether they come up with the same answers. Here we examined variability between four commonly used VBM pipelines in two large brain structural datasets. Spatial similarity and between-pipeline reproducibility of the processed gray matter brain maps were generally low between pipelines. Examination of sex-differences and age-related changes revealed considerable differences between the pipelines in terms of the specific regions identified. Machine learning-based multivariate analyses allowed accurate predictions of sex and age, however accuracy differed between pipelines. Our findings suggest that the choice of pipeline alone leads to considerable variability in brain structural markers which poses a serious challenge for reproducibility and interpretation. 1 The Center of Psychosomatic Medicine, Sichuan Provincial Center for Mental Health, Sichuan Provincial People’s Hospital, MOE Key Laboratory for Neuroinformation, High-Field Magnetic Resonance Imaging Key Laboratory of Sichuan Province, University of Electronic Science and Technology of China, Chengdu, China. 2 Institute of Brain and Psychological Sciences, Sichuan Normal University, Chengdu, China. 3 Neuroradiology Department, Fondazione Instituto Neurologico Carlo Besta, Milan, Italy. 4These authors contributed equally: Xinqi Zhou, Benjamin Becker. ✉email: ; COMMUNICATIONS BIOLOGY | (2022)5:913 | https://doi.org/10.1038/s42003-022-03880-1 | www.nature.com/commsbio 1 ARTICLE H COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-022-03880-1 uman fundamental and clinical neuroscience aims to determine the contribution of specific brain systems to mental processes and disorders, and neuroimaging approaches have been widely employed to this end. Due to its high spatial resolution and noninvasive nature, Magnetic Resonance Imaging (MRI)-based assessments of brain structure and function have become one of the most widely used neuroimaging techniques. However, the complexity and flexibility of workflows in MRI analyses, and differences between the handful of commonly used analysis software packages, may lead to high variability in neuroimaging results1. This variability challenges the interpretation of the results with respect to the precise mapping of mental processes and brain-based biomarkers for mental disorders. Compared to the processing of functional MRI (fMRI) data, brain morphometry analyses of T1-weighted structural images allow less processing variations and may have higher testretest reliability1–6. However, the choice of analytic software may still have a considerable impact on the results obtained. The variability in terms of whether and which specific brain regions pass the statistical threshold, in turn, impacts greatly on the interpretation of findings with respect to structure-function mapping or brain-based biomarkers and can significantly impede the sensitivity of subsequent neuroimaging meta-analyses. Neuroanatomical research has benefited tremendously from the development of automated computational approaches such as Voxel-based Morphometry (VBM), examining variations in regional gray matter volume, and the more recently developed surface-based approaches (e.g., examining cortical thickness). VBM represents one of the most commonly used brain structural analytic approaches to date (e.g., a simple literature search using the term “voxel-based morphometry” or “VBM” on PubMed revealed 6210 studies, https://pubmed.ncbi.nlm.nih.gov, from 1993 to November 19, 2020, see also publications for VBM and other approaches such as “cortical thickness” and “surface area” in PubMed depicted in Fig. S1). The standardized and highly automated VBM workflow includes segmentation of gray matter from other brain tissues, normalization into standard stereotactic space, and smoothing with a Gaussian kernel before inferential statistics are applied. The corresponding inferential voxel-wise statistical models commonly determine (1) between-group differences in regional gray matter volume (GMV), e.g., between patients and controls or men and women7–10, or (2) associations between individual variations in regional GMV and behavioral phenotypes, including learning, age, or disorder-relevant traits11–16. Significant differences or associations are commonly interpreted in a regional-specific fashion, e.g., mapping specific behavioral functions to specific brain systems, and determining which brain regions undergo age-related changes or which regions contribute to mental disorders. More recently, machinelearning-based multivariate analytic approaches such as Multivariate Pattern Analyses (MVPA) have been increasingly applied to VBM data to detect subtle and spatially distributed patterns of brain structural variations to improve biomarker-based diagnostics of mental disorders17–19. MVPA aims at determining variations in the spatial pattern across multiple voxels simultaneously and is thus often more sensitive in detecting betweengroup differences or brain structural associations. The approach is based on training pattern recognition algorithms, for example, brain structural data, and can be applied to new data to predict group membership (e.g., patients vs. controls, or women vs. men) or individual variations in a continuous variable such as age. A number of software packages have been developed and are widely utilized for VBM analyses. Among them, the currently most widely used ones are the Computational Anatomy Toolbox (CAT, www.neuro.uni-jena.de/cat), which is implemented in the Statistical Parametric Mapping software (SPM, https://www.fil. 2 ion.ucl.ac.uk/spm/software/spm12/), and FSLVBM and FSLANAT, which are based on the FMRIB Software Library (FSL, https://fsl.fmrib.ox.ac.uk). To enhance the robustness and reproducibility of neuroimaging analyses, new modular preprocessing pipelines for structural MRI (e.g., sMRIPrep, https://www. nipreps.org/smriprep/) have been recently developed. Although the software packages generally employ similar processing steps to volumetric T1-weighted (anatomical) MRI data, differences in specific processing steps and their implementation exist. This raises the question of whether the choice of specific software and the application of software-specific default processing configuration (...truncated)