Partial volume effect as a hidden covariate in DTI analyses
Research Highlights
► In fiber tract simulations, DTI measures (e.g., FA, MD) correlate with bundle volume. ► Confirming these simulations, FA was correlated with bundle volume in the cingulum. ► Correlations of FA with gender change when bundle volume is included as a covariate. ► We suggest to include tract volume as a covariate to improve DTI analysis specificity.
Introduction
Diffusion tensor imaging (DTI) is a non-invasive imaging technique that can provide information about brain microstructure and the directional organization of neural fiber tissue in vivo by measuring the self-diffusion of water molecules (Basser et al., 1994). In brain white matter (WM), diffusion of water is less hindered along than perpendicular to axons, making the local diffusion dependent on local microstructure (Beaulieu, 2002). DTI was first used clinically in schizophrenia (Buchsbaum et al., 1998) and leukoaraiosis (Jones et al., 1999b), where regional changes in diffusion anisotropy or trace were observed in patients but not in healthy controls. These diffusion changes suggest a structural change, which could be detected more easily on DTI scans than on conventional MR images. Since then, the use of DTI in both fundamental research and clinical studies has exploded, with almost a third of all studies on DTI discussing the development of fiber tractography (FT) methods, e.g., Mori et al. (1999), or the use of FT in DTI analyses. FT was initially applied as a method to investigate “brain connectivity” (Basser et al., 2000) and is now often used to increase specificity of observed radiological findings with respect to patient disability, for instance in multiple sclerosis (Wilson et al., 2003, Ciccarelli et al., 2008).
In recent years, DTI and FT have been used extensively to study the microstructural properties of WM fiber pathways. The developing/aging brain, in particular, has been the research topic of many investigations,1 with studies including subjects ranging from neonates to aging adults. It has been shown repeatedly that the FA of several WM regions (e.g., the cingulum bundles and the uncinate fasciculi) increases during maturation and subsequently decreases with age above the age of approximately 30 years (Bastin et al., 2010, Hasan et al., 2009, Hsu et al., 2008, Hsu et al., 2010, Jones et al., 2006, Lebel et al., 2008, Lebel et al., 2010, Sala et al., in Press, Salat et al., 2005, Voineskos et al., 2010, Voineskos et al., in Press). Most of these studies also show an inverse relationship (decrease followed by increase) for the mean diffusivity (MD) and link these diffusion changes to differences in microstructural organization within the WM (Dubois et al., 2008, Lebel et al., 2008). More specifically, changes in radial diffusivity (RD, diffusivity perpendicular to the predominant diffusion direction) and axial diffusivity (AD, diffusivity along the predominant diffusion direction) are believed to reflect different microstructural processes in the WM (Pierpaoli et al., 2001, Song et al., 2002, Song et al., 2003).
A particular aspect that is known to affect the accuracy of estimated DTI metrics – but which is not always considered a potential cause for correlations or differences in quantitative diffusion analyses – is the partial volume effect (PVE). Reflecting the intra-voxel heterogeneity of different tissue organizations (Alexander et al., 2001, Frank, 2001, Oouchi et al., 2007), Alexander et al. (2001) mentioned that “the PVE could cause diffusion-based characterization of cerebral ischemia and white matter connectivity to be incorrect”. Pfefferbaum and Sullivan (2003) have shown that the PVE is also present in the calculation of diffusion measures when averaging data values over regions of interest (ROIs). WM segmentation and semi-automated ROI delineation were used to outline the genu and splenium of the corpus callosum (CC), yielding increased MD values compared to the MD at the center of the WM bundles. This indicates a contamination of the outer WM voxels with its surrounding tissue, which, for the midsagittal genu and splenium of the CC, consists mostly of cerebrospinal fluid (CSF). Several options to mitigate such a CSF contamination have been proposed, such as CSF suppression using fluid-attenuated inversion recovery (FLAIR) acquisition sequences (Papadakis et al., 2002, Cheng and et al.., in Press), or using a two-tensor model (Pierpaoli and Jones, 2004, Pasternak et al., 2009) to remove the CSF contamination during tensor estimation. However, most DTI studies use neither of these techniques, which leave PVEs with CSF a relevant issue. As partial voluming is not only between WM bundles and CSF, but also, for instance, between different WM bundles, investigations into the effects of the PVE are important to improve quantitative diffusion analyses.
There are several confounding factors related to the PVE that may affect DTI metrics indirectly. For instance, as total WM volume changes with age, and therefore the thickness of some fiber bundles, the relative contribution of PVE-contaminated voxels will be different between bundles of different size (thicker bundles will have a lower contribution of PVE voxels to the entire bundle than thinner bundles), which may introduce a bias in the estimated measures (see Fig. 1 for a schematic example). Not only is bundle volume potentially a hidden covariate in the analysis of DTI metrics, but the orientation and curvature of a bundle may also alter the PVE and thus the diffusion measures.
In this work, we hypothesize that hidden covariates, such as bundle thickness (in the following also referred to as “volume”, assuming a constant bundle length and cross-sectional shape), orientation, curvature, and shape modulate the PVE intrinsically and, subsequently, affect the estimated DTI metrics. Previous studies show support for this hypothesis. For instance, investigations of brain volume changes with age show a decline in total WM volume from around the age of 30 (Courchesne et al., 2000, Liu et al., 2003, Resnick et al., 2003), matching the age-FA relation mentioned previously. Another study shows a left-sided co-lateralization of FA and concomitant bundle volume, potentially indicative of a more general agreement between morphometry and diffusion properties (Huster et al., 2009).
Using simulations of synthetic diffusion phantoms (Leemans et al., 2005) we determine whether the PVE-related covariates (volume, orientation, and curvature) affect the estimated diffusion measures. With these simulations, it is possible to change the volume, orientation, or curvature of a bundle independently while keeping all other configurational properties fixed. This allows for investigations of only the specific covariate of interest in relation to the estimated DTI metrics. In addition, the cingulum bundles and the CC of 55 healthy subjects are reconstructed using FT to examine whether the PVE-related confounding factors are present in experimental DTI data. The interest by many researchers in these bundles has resulted in an abundance of information about diffusion changes, and thus valuable reference material for this study (Davis et al., 2009, Huster et al., 2009, Jones et al., 2006, Lebel et al., 2008, Malykhin et al., 2008, Salat et al., 2005). The cingulum does not interface with CSF-filled spaces, in contrast with the CC, which is partially adjacent to the lateral ventricles and the longitudinal fissure. In regions where there is proximity to the ventricles, for example, one observes “spikes” in the MD values (Jones et al., 2005). The large difference in the surrounding tissues makes these bundles ideal candidates to determine the potential effect of PVE-related covariates on the different diffusion parameters. By investigating correlations between the volume of the bundles and specific diffusion properties of these bundles, the presence of such a hidden covariate may be revealed.
Our results demonstrate that DTI metrics are indeed correlated with volume, orientation, and curvature of a fiber bundle. As such, several conclusions drawn from previous analyses – aging studies in particular – should be nuanced in the light of these PVE-related covariates in order to correctly classify whether the observed changes in diffusion measures originate from either changes in macrostructural/morphological or microstructural properties, or a combination of both. Observed relations between age and diffusion properties are altered by the inclusion of volume as a covariate, which indicates that it is required to include this confound in quantitative analyses. Preliminary results of this work on PVE-related covariates have been presented at the 2010 Joint ISMRM–ESMRMB meeting in Stockholm, Sweden (Vos et al., 2010).
Section snippets
Fiber bundle simulations
Simulations of neural fiber bundles were performed according to Leemans et al. (2005) to investigate the following potentially confounding factors in FT-based analyses: (i) fiber bundle thickness (predefined range: 9–13 mm); (ii) pathway orientation (in-plane rotation range: 0–15°); and (iii) bundle curvature (inverse radius range: 0.035–0.055 mm− 1). Keeping all other simulation parameters unchanged, only the contribution of PVE-contaminated voxels to the bundles differs between the simulations.
PVE-related covariates in simulations
All six sets of simulations (changes in volume, orientation, and curvature, for both the cingulum and CC simulations) showed a clear effect of these PVE-related covariates on the estimated DTI metrics. Bundle volume demonstrated a monotonous relation with the diffusion measures (Fig. 5). By contrast, the other PVE-modulating factors generally displayed a high degree of non-monotonicity (Fig. 6, Fig. 7).
Experimental data
Correlations between segment volume and DTI metrics have been visualized in Fig. 8 for the
Discussion
Several factors, e.g., bundle thickness, orientation, and curvature may change the PVE and thus the analysis and estimation of bundle-averaged DTI metrics. These PVEs originate from the acquisition: signal averaging over finite-size voxels may include more than one structure. As already illustrated in Fig. 1, bundles can be influenced differently by the PVE depending on bundle thickness. However, a single bundle may also be affected by its position relative to the acquisition matrix. Consider
Acknowledgments
This work was financially supported by the Care4Me (Cooperative Advanced REsearch for Medical Efficiency) pan-European research program ITEA (Information Technology for European Advancement). The authors would like to thank Dr. John Evans, Chief Physicist of CUBRIC, for assistance in acquiring the MR data.
References (69)
- et al.
MR diffusion tensor spectroscopy and imaging
Biophys. J.
(1994) - et al.
Quantifying the effects of normal ageing on white matter structure using unsupervised tract shape modelling
Neuroimage
(2010) - et al.
Diffusion tensor imaging in children and adolescents: reproducibility, hemispheric, and age-related differences
Neuroimage
(2007) - et al.
Diffusion-based tractography in neurological disorders: concepts, applications, and future developments
Lancet Neurol.
(2008) - et al.
Assessing the effects of age on long white matter tracts using diffusion tensor tractography
Neuroimage
(2009) - et al.
The relationship between diffusion tensor imaging and volumetry as measures of white matter properties
Neuroimage
(2008) - et al.
Development and aging of the healthy human brain uncinate fasciculus across the lifespan using diffusion tensor tractography
Brain Res.
(2009) - et al.
Gender differences and age-related white matter changes of the human brain: a diffusion tensor imaging study
Neuroimage
(2008) - et al.
Microstructural white matter changes in normal aging: a diffusion tensor imaging study with higher-order polynomial regression models
Neuroimage
(2010) - et al.
White and gray matter development in human fetal, newborn and pediatric brains
Neuroimage
(2006)