Elsevier

NeuroImage

Volume 30, Issue 2, 1 April 2006, Pages 436-443
NeuroImage

Reliability in multi-site structural MRI studies: Effects of gradient non-linearity correction on phantom and human data

https://doi.org/10.1016/j.neuroimage.2005.09.046Get rights and content

Abstract

Longitudinal and multi-site clinical studies create the imperative to characterize and correct technological sources of variance that limit image reproducibility in high-resolution structural MRI studies, thus facilitating precise, quantitative, platform-independent, multi-site evaluation. In this work, we investigated the effects that imaging gradient non-linearity have on reproducibility of multi-site human MRI. We applied an image distortion correction method based on spherical harmonics description of the gradients and verified the accuracy of the method using phantom data. The correction method was then applied to the brain image data from a group of subjects scanned twice at multiple sites having different 1.5 T platforms. Within-site and across-site variability of the image data was assessed by evaluating voxel-based image intensity reproducibility. The image intensity reproducibility of the human brain data was significantly improved with distortion correction, suggesting that this method may offer improved reproducibility in morphometry studies. We provide the source code for the gradient distortion algorithm together with the phantom data.

Introduction

Multi-site and longitudinal neuroimaging studies are increasingly becoming a standard element of clinical neuropsychiatric research for diagnosing and evaluating neurological impairments (Ashburner et al., 2003, Grundman et al., 2002, Fox and Schott, 2004). One of the challenges of both longitudinal and multi-site studies is to minimize image variability caused by technological factors (e.g., hardware differences, hardware imperfections), as such variability may be confounded with specific disease-related changes in the images thereby limiting the power to detect and follow the progression of disease biomarkers. Optimization of image reproducibility motivates the calibration of acquisition protocols and the characterization and correction of scanner-specific image variability effects. This is particularly important when data from multiple sites and MRI vendors are to be combined.

An important task in this effort is to correct for site-specific image distortions in order to allow accurate cross-site comparisons of quantitative morphometry results. Image distortions can potentially affect the accuracy of volume (Fischl et al., 2002), shape (Miller, 2004) and boundary (Barnes et al., 2004) measurements. Although distortions in MRI can arise from several factors, one of the most prominent in structural MRI is imaging gradient non-linearity, which degrades both geometric and image intensity accuracy. While in principle gradient distortions may be addressable using manufacturer-supplied software, the currently available correction algorithms work only in two-dimension (2-D) providing an incomplete solution to the problem (Wang et al., 2004a). Three-dimensional (3-D) algorithms to correct gradient non-linearity distortions have been investigated using phantoms. To summarize, two main correction methods have been developed: (a) 3-D measurement of the geometric displacements due to distortions using specially designed phantoms followed by an image transformation to perform the correction (Wang et al., 2004b, Wang et al., 2004c, Langlois et al., 1999) and (b) 3-D calculation of the geometric displacements from the spherical harmonic expansion for the representation of the magnetic fields generated by the gradient coils (Schmitt, 1985, Janke et al., 2004, Wald et al., 2001). The second method was used in this work. As yet, there is no quantitative study that systematically compares these correction methods. More importantly, no study investigates the effects of these distortion correction methods on test–retest reproducibility of multi-site human structural MRI data.

The purpose of this work was (i) to quantitatively characterize and correct site-specific image distortions caused by gradient non-linearity in a phantom study, and (ii) to assess if gradient non-linearity distortion correction improves image reproducibility when the same subjects are scanned at multiple sites in multiple sessions. To keep our results independent of brain morphometry analysis tools, here, we focus only on the reproducibility of image intensity for the human data. Parts of these results have been presented at recent meetings (Jovicich et al., 2004, Jovicich et al., 2003).

Section snippets

Human and phantom image data acquisitions

Four sites with clinical 1.5 T whole body scanners used in regular functional and structural MRI studies participated in this study. These systems included: (a) GE Medical Systems with Cardiac Resonator Module (CRM) gradient coils (maximum strength = 40 mT/m, slew rate = 150 T/m/s) at Duke University Medical Center (Duke) and at Brigham and Women's Hospital (BWH); (b) GE Medical Systems with Brain Resonator Module (BRM) gradient coils at the University of California San Diego (UCSD, 22 mT/m,

(1) Gradient distortion correction: phantom validation

To validate the distortion correction we measured the diameter of the phantom images obtained at four sites, and compared the measures with the true diameter, with and without image distortion correction. Fig. 1 gives a schematic summary of the gradient distortion correction process, with sample phantom images from the MGH site. The spherical harmonic coefficients from the MR vendor's gradient were used to calculate 3-D displacement and intensity correction fields within the scanner's field of

Conclusions

Structural MRI studies offer the potential for quantifying subtle brain structural differences between patient populations or changes over time, such as during development or atrophy in neurodegenerative diseases. Further, the ability to combine calibrated data acquired across multiple sites offers the possibility for increased statistical power by analyzing large collections of datasets. However, the effectiveness of such approaches is limited by image reproducibility errors, which can

Acknowledgments

The authors thank: (a) our human phantoms, (b) Steve Pieper and Charles Guttmann (Brigham Women's Hospital) for coordinating the phantom scans at BWH, (c) Gabriela Czanner (Massachusetts General Hospital) for her support on statistical analyses and (d) Gary Glover (Stanford University), Larry Frank (University of California, San Diego) and Jason Polzin (GE Healthcare Technologies), Eva Eberlein (Siemens Medical Solutions, Erlangen, Germany), for their support on making available the gradient's

Cited by (0)

View full text