ComBatLS: A location- and scale-preserving method for multi-site image harmonization

1889 

Abstract Submission 

Margaret Gardner¹, Russell Shinohara¹, Richard Bethlehem², Rafael Romero-García³, Varun Warrier⁴, Sheila Shanmugan¹, Jakob Seidlitz¹, Aaron Alexander-Bloch¹, Andrew Chen⁵

¹University of Pennsylvania, Philadelphia, PA, ²Autism Research Centre, Department of Psychiatry, University of Cambridge, Cambridge, United Kingdom, ³University of Seville, Seville, Spain, ⁴University of Cambridge, Cambridge, Cambridgeshire, ⁵Medical University of South Caroline, Charleston, SC

Margaret Gardner  
University of Pennsylvania
Philadelphia, PA

Russell Shinohara  
University of Pennsylvania
Philadelphia, PA

Richard Bethlehem  
Autism Research Centre, Department of Psychiatry, University of Cambridge
Cambridge, United Kingdom

Rafael Romero-García  
University of Seville
Seville, Spain

Varun Warrier  
University of Cambridge
Cambridge, Cambridgeshire

Sheila Shanmugan  
University of Pennsylvania
Philadelphia, PA

Jakob Seidlitz  
University of Pennsylvania
Philadelphia, PA

Aaron Alexander-Bloch  
University of Pennsylvania
Philadelphia, PA

Andrew Chen  
Medical University of South Caroline
Charleston, SC

Recent work has leveraged massive datasets and advanced image harmonization algorithms to construct normative models of imaging-derived phenotypes (IDPs)[1,2]. These brain chart models, which can produce centile or z-scores to benchmark individuals' morphology within a population, are often fit on magnetic resonance imaging data collected across hundreds of scanners. One popular method for harmonizing these data is ComBat, which preserves the effects of specified covariates on the IDPs' means. However, evidence suggests that biological factors, such as sex, also impact an IDP's variance across a population[3]. These scale effects, which directly impact centile and z-score distributions, are not preserved by current harmonization methods. Thus, harmonization may induce error in centile and z-scores, particularly when factors that impact scale are distributed unequally across sites.

Here, we propose a new method in the ComBat family of harmonization tools, ComBatLS, that preserves biological variance in IDPs' location and scale. We tested ComBatLS's ability to preserve variation in scale and its impacts on centile and z-scores by harmonizing across sex-imbalanced artificial "sites" in data from the UK Biobank.

As in all ComBat versions, ComBatLS harmonizes across sites by targeting each IDP to a pooled mean while estimating and preserving the effects of designated covariates. However, prior ComBat models assumed that the variance of each IDP's distribution was consistent across sites. In ComBatLS, we incorporate a log-linear relationship between the error's standard deviation and the covariates[4], enabling the estimation and preservation of covariates' effects on IDP variance.

To test this new method, we utilized IDPs derived from the structural MRIs of 28619 participants in the UK Biobank (49.7% female, age 50-80 years), collected across 3 scanners using identical hardware and protocols[5]. T1-weighted structural images, along with T2 for ~98% of subjects, were processed using FreeSurfer v. 6 to extract global volumes for each tissue class (cortical gray matter, subcortical gray matter, white matter, and CSF), as well as surface area, thickness, and volume for each cortical region[6]. Subjects were randomly assigned to one of three simulated sites such that each site had a Male:Female ratio of 1:1, 1:4, and 4:1, respectively. We then harmonized IDPs across the simulated sites using 4 different ComBat configurations: the first preserved no covariate effects, while the remainder preserved the effects of sex and age estimated by a linear model[7], ComBat-GAM[8], and ComBatLS, respectively. Data harmonized by each configuration was then used to fit simple brain charts for each IDP using generalized additive models of location, scale, and shape.

We assessed ComBat configurations' influence by comparing sex effects in scale in each brain chart and, crucially, how the centile and z-scores derived from them were affected by each harmonization technique. Models fit on the complete, unharmonized data showed the true sex effect in scale in this dataset. Despite the confounding of sex- and site-effects in our simulations, ComBatLS accurately preserved the effects of sex on IDP variance (Fig 1). For both centile and z-scores, pairwise t-tests revealed that the magnitude of error for each subject's scores differed significantly across ComBat methods, with ComBatLS producing more accurate scores than the next most accurate method, ComBat-GAM (Fig 2), in 207 of 208 IDPs (pFDR centiles = <0.001 - 0.67, pFDR z-scores = <0.001 - 0.134). These results were stable across 10 permutations of subjects' site assignments.

We propose ComBatLS as a robust method for harmonizing neuroimaging data across sites while preserving biologically meaningful differences in scale for accurate centile and z-score estimation. ComBatLS is available alongside other ComBat harmonization tools at https://github.com/andy1764/ComBatFamily.

Methods Development ¹

Neuroanatomy Other ²

Informatics

Modeling

MRI

Open-Source Code

Sexual Dimorphism

Statistical Methods

STRUCTURAL MRI

Other - multisite harmonization