henhao
Sign in
NeuroscienceNeuroImage
Peer-reviewed journal article

Harmonization of cortical thickness measurements across scanners and sites

Jean‐Philippe FortiniD, Nicholas C. CulleniD, Yvette I. ShelineiD, Warren D. TayloriD, Irem AselciogluiD, Philip A. CookiD, Phil Adams, Crystal M. CooperiD, Maurizio Fava M. DiD, Patrick J. McGrathiD, Melvin G. McInnisiD, Mary Louise PhillipsiD, Madhukar H. TrivediiD, Myrna M. WeissmaniD, Russell Taki ShinoharaiD

Abstract

With the proliferation of multi-site neuroimaging studies, there is a greater need for handling non-biological variance introduced by differences in MRI scanners and acquisition protocols. Such unwanted sources of variation, which we refer to as \"scanner effects\", can hinder the detection of imaging features associated with clinical covariates of interest and cause spurious findings. In this paper, we investigate scanner effects in two large multi-site studies on cortical thickness measurements, across a total of 11 scanners. We propose a set of general tools for visualizing and identifying scanner effects that are generalizable to other modalities. We then propose to use ComBat, a technique adopted from the genomics literature and recently applied to diffusion tensor imaging data, to combine and harmonize cortical thickness values across scanners. We show that ComBat removes unwanted sources of scan variability while simultaneously increasing the power and reproducibility of subsequent statistical analyses. We also show that ComBat is useful for combining imaging data with the goal of studying life-span trajectories in the brain.

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
1,585
FWCI
28.10
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (10 Jun 2017) · CC BY 4.0

ComBat, an empirical Bayes location-and-scale model from genomics, removes additive and multiplicative scanner effects from regional cortical thickness across 11 scanners while preserving age associations.

Multi-site MRI studies add scanner- and protocol-driven variation that can hide clinical associations or produce spurious findings, and standard confound adjustment for outcome prediction was reported to do no better than ignoring confounding. The work extends ComBat, previously applied to DTI, to cortical thickness and adds general diagnostics for detecting scanner effects.

Key results

  1. ShownRaw cortical thickness carried strong scanner effects in both cohorts. Site differed significantly in both mean and variance of median thickness, and many individual regions of interest (ROIs) were associated with site.

    EMBARC: ANOVA on median thickness p = 1.1 × 10^−10; Bartlett p = 1.8 x 10^−7; 53 ROIs (Bonferroni) or 60 ROIs (maxT, B = 10,000) site-associated; 41 ROIs with site-dependent variance. VDLC: ANOVA p = 2.2 × 10^−16; 86 or 87 ROIs; Bartlett p = 0.0013; 3T and 1.5T scanners differed. · 3.1 Visualization of site effects; 3.2 Quantification of site effects in the EMBARC…; A.1 Quantification of site effects in the VDLC…

  2. ShownOnly ComBat removed site-specific differences in variance (scale effects). After all three harmonizations, no individual ROI remained significantly associated with site under Bonferroni or maxT adjustment.

    Bartlett test after harmonization. EMBARC: ComBat p = 0.42 vs Residuals p = 2.53 x 10^−8 and Adjusted Residuals p = 3.08 x 10^−8. VDLC: ComBat p = 0.46 vs 0.03 and 0.01. EMBARC median-thickness ANOVA: p = 0.997 (Residuals), 0.0498 (Adjusted), 0.0473 (ComBat). · 3.3 Removal of site effects with harmonization; A.2 Removal of scanner effects in the VDLC study…

  3. ShownComBat gave the lowest site-prediction accuracy in both cohorts. In EMBARC, support vector machine (SVM) classifiers on ComBat-harmonized features were near permutation chance, which the authors suggest may reflect ComBat also correcting scale effects.

    Radial-kernel SVM, B = 10,000 repeats of 10-fold CV. EMBARC accuracy: raw 76.6%, Residuals 40.5%, Adjusted Residuals 38.7%, ComBat 36.3% (permutation chance 36.9%). VDLC: raw 67.7%, Residuals 43.4%, Adjusted Residuals 44.4%, ComBat 41.0%. · 3.3 Removal of site effects with harmonization; A.2 Removal of scanner effects in the VDLC study…

  4. ShownIn EMBARC, the covariate-aware methods (Adjusted Residuals and ComBat) strengthened the age association and improved linear age prediction. Unadjusted residualization added little age signal and predicted age poorly, which the authors attribute to age being confounded with site.

    Variance in median thickness explained by age: raw 23%, Residuals 26%, Adjusted Residuals and ComBat 33%. Age RMSE over B = 1000 random k-fold splits (k ∈ {3,5,10}): both methods improved on raw with linear regression, but ϵ-SVM regression gave little gain over raw. · 3.1 Visualization of site effects; 3.4 Associations with age

  5. ShownWhen the EMBARC and VDLC cohorts (different scanners and age ranges) were pooled, ComBat and Adjusted Residuals reduced variation between subjects and sharpened the decline of cortical thickness with age across the lifespan.

    Correlation of median cortical thickness with age in the pooled data (EMBARC 18-65 y, VDLC 58-95 y): raw −0.70, Residuals −0.26, Adjusted Residuals −0.77, ComBat −0.79. · 3.5 Life-span study by harmonizing the EMBARC…

Methods, data and reproducibility

Regional cortical thickness from the ANTs pipeline was harmonized with ComBat, a per-feature linear location-and-scale model with empirical Bayes shrinkage of additive and multiplicative site parameters. It was compared with unadjusted and covariate-adjusted OLS residualization using ANOVA, Bartlett tests, PCA/LDA, SVM site classification and age analyses.

  • EMBARC baseline structural MRI of unmedicated adults with major depression and healthy controls; 4 sites with 3T scanners from different vendors; 102 cortical ROIs · human · n = 187 after QC (200 MDD and 40 healthy recruited)
  • VDLC late-life depression study, last-visit structural MRI at Duke and Washington University in St-Louis on 7 scanners (1.5T and 3T); 98 cortical ROIs · human · n = 236 participants (117 LLD, 59 healthy controls)
  • Average labeled template built from a subset of Open Access Series of Imaging Studies (OASIS) participants, used to define cortical regions · human (OASIS participants)
Methods & tools
ANTs cortical thickness pipeline, ComBat (sva R package reference implementation), ComBatHarmonization (R/Matlab), neuroCombat (Python), Residuals and Adjusted Residuals OLS harmonization, Principal component analysis (PCA), Linear discriminant analysis (LDA), Support vector machine (radial basis kernel), ϵ-SVM regression (SVR), ANOVA with Tukey correction and Bartlett test, One-step maxT permutation procedure
Code
https://github.com/Jfortin1/ComBatHarmonizationhttps://github.com/ncullen93/neuroCombat
Software & environment
R version 3.2.0 for all postprocessing; ComBat reference implementation from the R sva package; the adapted harmonization code is provided in R, Matlab and Python.
Reproducibility gaps
The text gives no repository, accession or access procedure for the EMBARC or VDLC images. Acquisition parameters are deferred to tables and prior publications that are not reproduced in the text.
Limitations (5)
  • ComBat's performance when site is confounded with covariates still needs improvement. The authors plan to add an inverse probability weighting scheme to address this. · stated by the authors
  • The ComBat used here is cross-sectional. The authors plan a time-dependent version for subjects scanned at several time points on different scanners or with changed scanning parameters, and only the last-visit scan of each VDLC participant was analyzed. · stated by the authors
  • All the harmonization methods compared are linear models, so they may suit linear prediction but not nonlinear ϵ-SVM regression. With ϵ-SVM, age prediction barely improved over the raw data. · stated by the authors
  • Preservation of biology was assessed only through age, which is a covariate that Adjusted Residuals and ComBat explicitly protect. Clinical scores such as HAMD or QIDS were not tested for retained associations after harmonization. · noted in this summary
  • In EMBARC, the post-harmonization ANOVA p values for Adjusted Residuals (0.0498) and ComBat (0.0473) fall just under 0.05, yet these methods are described as corrected. Also, the reported VDLC subgroup counts (117 and 59) do not add up to the stated total of 236. · noted in this summary

Who should read:Read this if you pool structural MRI or other region-level imaging features across scanners or sites and want a tested ComBat harmonization workflow with diagnostics for detecting scanner effects.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBeer '20Sun '22Zhang '24Nielson '18B '26Ren '26
83% semantic matchNeuroscience
Joanne C. Beer, Nicholas James Tustison, Philip A. Cook+4 more·NeuroImage·Jul 5, 2020· 322 citations