Tract-specific anisotropy measurements in diffusion tensor imaging

Abstract

Diffusion tensor magnetic resonance imaging (DT-MRI) has been used to examine the microstructure of individual white matter tracts, often in neuropsychiatric conditions without identifiable focal pathology. However, the voxel-based group-mapping and region-of-interest (ROI) approaches used to analyse the data have inherent conceptual and practical difficulties. Taking the example of the genu of the corpus callosum in a sample of schizophrenic patients, we discuss the difficulties in attempting to replicate a voxel-based finding of reduced anisotropy using two ROI methods. Firstly we consider conventional ROIs; secondly, we present a novel tractography-based approach. The problems of both methods are explored, particularly of high variance and ROI definition. The potential benefits of the tractographic method for neuropsychiatric conditions with subtle and diffuse pathology are outlined.

Introduction

Diffusion tensor magnetic resonance imaging (DT-MRI) is a relatively new imaging technique (Basser et al., 1994) that has been rapidly adopted by neuropsychiatric researchers, though significant questions over its analysis remain unanswered. It uses a conventional MRI scanner but applies pulsed magnetic field gradients to encode the diffusivity of water in at least six non-collinear directions. It thereby acquires signals reflecting the orientational dependence of the apparent diffusivity of water, and from these measurements the bulk-averaged diffusion tensor can be calculated at each voxel (Basser et al., 1994). When diffusion is hindered by cellular architecture to a predominant direction, such as in white matter, it is anisotropic. In tissue where it has no predominant direction, such as in the ventricles, grey matter, or in a white matter voxel that includes crossing fiber populations, diffusion will appear more isotropic.

DT-MRI has yielded two significant advantages over other methods. Firstly, it provides quantitative measures of microstructural organization, such as fractional anisotropy (FA) (Basser and Pierpaoli, 1996), which have proved to be highly sensitive to white matter changes in a number of pathological conditions. Secondly, the orientational information can be used to resolve seemingly homogenous white matter into its constituent tracts (see, for example, Mori and van Zijl, 2002). Combined, these two advantages have engendered a change in the way white matter is considered. They have encouraged the application of DT-MRI to conditions where white matter pathology cannot be detected by other imaging modalities in vivo, such as cryptogenic epilepsy (Rugg-Gunn et al., 2001) and diffuse axonal injury (Arfanakis et al., 2002). Further, they have encouraged researchers to localise pathology to specific tracts — tracts that differ greatly in their connectivity, and therefore in their functionality. This is timely as neuropathology shifts from lesion models to models proposing that disconnectivity underlies symptoms (Marshall and Fink, 2003).

The methods of analysis used for DT-MRI data are primarily whole-brain voxel-based analysis (VBA) and region-of-interest (ROI) approaches. The promise of VBA is that it should be operator-independent, and allows consideration of entire brain volumes. This can be very useful where changes are diffuse or an a priori hypothesis regarding their location is unavailable. However, adapting the voxel-based approach, developed for functional and structural imaging data, to DT-MRI data is not necessarily straightforward. First, co-registration of low-resolution, high-contrast FA maps may generate significant mis-registration and partial volume artefacts in regions of high and low anisotropy, for example, around the ventricles. Second, the accurate localisation of differences to specific tracts is difficult with VBA. The data are often heavily smoothed as part of the pre-processing to ensure that the ensuing statistical approach used is valid (Ashburner and Friston, 2000). This results in inherently low resolution parametric maps from which to infer group differences and also sensitizes the detection of group differences to the size of the smoothing filter used (Jones et al., 2005). Three-dimensional clusters of voxels exhibiting significant group differences will not typically lie neatly within a single tract, and resolving a cluster into component tracts by reference to anatomical atlases is compromised by both by the limited resolution of the parametric maps and by the limited white matter detail such atlases contain.

The majority of DT-MRI studies have pursued ROI-based analyses, either in their own right, or to confirm VBA findings. ROIs offer a powerful way of directing attention to hypothesis-specific tracts implicated in a disorder. They have difficulties of their own, however, as we shall discuss in this article.

Schizophrenia is a condition that has seemed highly suitable for study with DT-MRI. There is evidence of white matter pathology that is ill-defined at the gross anatomical level, but may underpin disconnectivity models of schizophrenia (David, 1994, Friston and Frith, 1995, Bullmore et al., 1997, Andreasen et al., 1998, Crow, 1998). A review of DT-MRI studies in schizophrenia (Kanaan et al., in press) reveals considerable variation in the results obtained to date, however. While this may partly be due to epidemiological issues of sample size and group matching, it may also be in part due to the variation in regions examined and choice of analytic strategy. Most of the studies have used conventional ROIs, placed either on the axial FA or T2 reference images, and most did not choose the same regions. However, even when the regions were similar, as in (Foong et al., 2000, Sun et al., 2003, Kumra et al., 2004), results were still conflicting. This raises the question: Just how suitable is the ROI method for the demonstration of hypothesised abnormalities in DT-MRI in schizophrenia?

We explore this in the context of a DT-MRI dataset from patients with schizophrenia and healthy controls. A previous whole-brain voxel-based analysis of the data set (Kanaan et al., 2004) found statistically significant differences in FA in several areas, including the genu of the corpus callosum (Fig. 1); however, concerns that these might represent registration errors prompted us to validate these results with a method that does not require registration. The present study thus assesses the validity of those findings using two methods of directly examining FA in the genu and examines the limitations of these approaches. Firstly we consider conventional ROIs, before presenting a novel tractography-based ROI technique (Jones et al., in press). Tractography uses the directional information in DT-MRI data to generate three-dimensional reconstructions of white matter pathways (Basser et al., 2000). While DT-MRI tractography has produced striking images, it has suffered from a number of problems, for example: reliability (within the same data-set) is operator-dependent on seedpoint placement; assessment of validity is hampered by the lack of a neuroanatomical white matter ‘gold standard’; and there are difficulties in deterministically resolving the crossing or meeting of different white matter bundles. Though tractography has found use in anatomical exploration (Catani et al., 2003) and tumour definition (Mori et al., 2002a), we employ it here to define our ROI for further analysis.