A framework for the analysis of phantom data in multicenter diffusion tensor imaging studies

Abstract

Diffusion tensor imaging (DTI) is commonly used for studies of the human brain due to its inherent sensitivity to the microstructural architecture of white matter. To increase sampling diversity, it is often desirable to perform multicenter studies. However, it is likely that the variability of acquired data will be greater in multicenter studies than in single-center studies due to the added confound of differences between sites. Therefore, careful characterization of the contributions to variance in a multicenter study is extremely important for meaningful pooling of data from multiple sites. We propose a two-step analysis framework for first identifying outlier datasets, followed by a parametric variance analysis for identification of intersite and intrasite contributions to total variance. This framework is then applied to phantom data from the NIH MRI study of normal brain development (PedsMRI). Our results suggest that initial outlier identification is extremely important for accurate assessment of intersite and intrasite variability, as well as for early identification of problems with data acquisition. We recommend the use of the presented framework at frequent intervals during the data acquisition phase of multicenter DTI studies, which will allow investigators to identify and solve problems as they occur.

Keywords: DTI; accuracy; diffusion tensor imaging; multicenter; pediatric; reproducibility.

Copyright © 2012 Wiley Periodicals, Inc.

Figures

Figure 1

Schematic of the first step of the analysis framework: outlier identification. A median map is calculated for the desired tensor‐derived metric (for example, FA). Each individual time point FA image is subtracted from the median to identify outlier datasets.

Figure 2

Sample difference from median maps for the simulated data validation tests for outlier identification. Bias and noise levels are indicated below the affected images. The outlier identification step clearly identifies data sets with both a bias and low SNR.

Figure 3

Intersite and intrasite variability maps of FA and Trace(D) for simulated data.

Figure 4

ICC maps for simulated data. No site difference is indicated when inter and intrasite ICC are equal (ICCinter = ICCintra = 0.5).

Figure 5

Difference from median maps for the PedsMRI physical phantom. Site 3 clearly shows a significant site difference compared to Sites 2 and 5.

Figure 6

Inter‐ and intrasite variability maps for the PedsMRI physical phantom. Intersite variability is clearly greater than intrasite variability, indicating a significant site difference.

Figure 7

Difference from median maps for the PedsMRI living phantom data. Datasets identified as outliers are marked by white, solid and hashed black circles.

Figure 8

Inter‐ and intrasite variability maps for the PedsMRI living phantom data. Intersite variability is greater than intrasite variability. This is strongest at csf/tissue and white matter/gray matter interfaces, indicating significant misregistration between sites.

Figure 9

Example of a ghosting artifact that resulted in an outlier dataset from the PedsMRI living phantom data.