Exploration of scanning effects in multi-site structural MRI studies

Abstract

Background: Pooling of multi-site MRI data is often necessary when a large cohort is desired. However, different scanning platforms can introduce systematic differences which confound true effects of interest. One may reduce multi-site bias by calibrating pivotal scanning parameters, or include them as covariates to improve the data integrity.

New method: In the present study we use a source-based morphometry (SBM) model to explore scanning effects in multi-site sMRI studies and develop a data-driven correction. Specifically, independent components are extracted from the data and investigated for associations with scanning parameters to assess the influence. The identified scanning-related components can be eliminated from the original data for correction.

Results: A small set of SBM components captured most of the variance associated with the scanning differences. In a dataset of 1460 healthy subjects, pronounced and independent scanning effects were observed in brainstem and thalamus, associated with magnetic field strength-inversion time and RF-receiving coil. A second study with 110 schizophrenia patients and 124 healthy controls demonstrated that scanning effects can be effectively corrected with the SBM approach.

Comparison with existing method(s): Both SBM and GLM correction appeared to effectively eliminate the scanning effects. Meanwhile, the SBM-corrected data yielded a more significant patient versus control group difference and less questionable findings.

Conclusions: It is important to calibrate scanning settings and completely examine individual parameters for the control of confounding effects in multi-site sMRI studies. Both GLM and SBM correction can reduce scanning effects, though SBM's data-driven nature provides additional flexibility and is better able to handle collinear effects.

Keywords: ICA; Multi-site; Multivariate; SBM; sMRI.

Fig. 1

A flow chart of the SBM model. (a) ICA model; (b) pairwise association analyses; (c) estimation of threshold p-value. Red lines L1 and L2 represent the linear fittings to the two segments of the component curve (the blue dotted curve); O1 denotes the intersection of L1 and L2; the green line L3 represents the line connecting the origin and the intersection O1; O2 denotes the intersection of L3 and the component curve based on which threshold p-value is determined. (For interpretation of the references to color in figure legend, the reader is referred to the web version of the article.)

Fig. 2

Spatial maps of the scanning-related components identified in the BIG data (|Z| > 2).

Fig. 3

Boxplots of two components exhibiting the most significant scanning effects in the BIG data; (a) IC9 loadings versus magnetic field strength-inversion time-pixel bandwidth; (b) IC7 loadings versus receiving coil.

Fig. 4

Spatial maps of the scanning-related components identified in the MCIC data (|Z| > 2).

Fig. 5

Boxplots of the component exhibiting the most significant scanning effect in the MCIC data; (a) IC53 loadings versus magnetic field strength; (b) IC53 loadings versus TR/TE.

Fig. 6

A spatial map of the SZ-related voxels identified with VBM (FDR control) in the uncorrected or GLM-corrected, but not in the SBM-corrected data.

Fig. 7

Spatial maps (|Z| > 2) and boxplots of the most significant SZ-discriminating components identified in the uncorrected (IC18), GLM-corrected (IC20) and SBM-corrected (IC2) data after smoothing. All the components are spatially consistent, highlighting the frontal temporal network.

Fig. 8

Spatial maps of questionable SZ-discriminating components identified in the uncorrected and GLM-corrected data after smoothing.