Diffusion tensor imaging segmentation of white matter structures using a Reproducible Objective Quantification Scheme (ROQS)

Abstract

Reproducible Objective Quantification Scheme (ROQS) is a novel method for regional white matter measurements of diffusion tensor imaging (DTI) parameters that overcomes the limitations of previous approaches for analyzing large cohorts of subjects reliably. ROQS is a semi-automated technique that exploits the fiber orientation information from the diffusion tensor in conjunction with a binary masking and chain-linking algorithm to segment anatomically distinct white matter tracts for subsequent quantitative analysis of DTI parameters such as fractional anisotropy and apparent diffusion coefficient. When applied to 3-T whole-brain DTI of normal adult volunteers, ROQS is shown to segment the corpus callosum much faster than manual region of interest (ROI) delineation, and with better reproducibility and accuracy.

Introduction

Diffusion tensor magnetic resonance imaging (DTI) has expanded the boundaries of diagnostic imaging by examining the diffusion of water in brain tissue (Basser et al., 1994, Pierpaoli et al., 1996). Recently, studies have shown that DTI can be used to provide diagnosis of disease conditions in cerebral ischemia, acute stroke, and multiple sclerosis (Virta et al., 1999, Mukherjee et al., 2000, Werring et al., 1999). In white matter, water diffuses more readily along the orientation of axonal fibers than across the fibers due to hindrance from structural elements such as the myelin sheath. The apparent diffusion coefficient (D_av) is a rotationally invariant measure of the magnitude of diffusion. The degree of directionality of diffusion is termed anisotropy, and can be measured as the variation in the eigenvalues of the diffusion tensor (Basser et al., 1994). An important observation is that these anisotropy measures can be obtained without diagonalization of the diffusion tensor to calculate the eigenvalues, but via scaled invariants (Ulug and van Zijl, 1999). Fractional anisotropy (FA) is sensitive to changes in white matter integrity (Watts et al., 2003). FA loss and D_av rise have been demonstrated in a number of traumatic brain injury (TBI) studies (Huisman et al., 2004, Lee et al., 2003, Huisman et al., 2003, Arfanakis et al., 2002). It is reasonable to assume that these quantitative characteristics are indicative of white matter damage in a variety of pathologies.

The quantification of DTI for investigating white matter abnormalities is generally approached using one of two methods: voxel-based analysis (VBA) where spatially normalized data sets are compared at a voxel by voxel scale or region of interest (ROI) analysis in which a specified region is measured directly. In cases where it is impractical to predict anatomical domains of damage a priori, researchers tend to use a voxel-based approach to characterize statistical differences between groups (Eriksson et al., 2001, Rugg-Gunn et al., 2001). VBA involves spatial normalization of brain images to a stereotactic 3D space (Friston, 1995, Mazziotta et al., 1995, Thompson and Toga, 1997, Grenander and Miller, 1998). In order to produce a more normalized distribution of image data, a smoothing function is typically applied to the images (Ashburner and Friston, 2000, Good et al., 2001, Watkins et al., 2001). Statistical differences between groups are then made on a voxel by voxel basis to determine variations in tissue composition.

While a voxel-based strategy has the advantage of evaluating the entire brain in a model-free manner, and is therefore suitable for the identification of unexpected areas of white matter pathology, there are several limitations. A central disadvantage of this approach is that differences in gross anatomical morphology among subjects may influence spatial normalization and thus artificially inflate measurement differences. Moreover, many normalization algorithms use smoothing functions which introduce “blur” into the image, and violate the original uniformity of voxel size present in the original images, thus creating noise in the measurement (Jones et al., 2005). In other words, smoothing algorithms can alter the voxel size present in the original image due to low-pass filters. It has been shown that the size of the smoothing kernel can dramatically affect the analysis of the data (Jones et al., 2005). Researchers have suggested the possibility that VBA has reduced and inconsistent sensitivity in specific regions of the brain, especially those with greater anatomical variability (Quarantelli et al., 2002, Tisserand et al., 2002, Ciccarelli et al., 2003).

Most researchers use ROI analysis when it is possible to hypothesize specific areas of the brain that are implicated in disease. ROI analysis normally requires manual tracing of readily identifiable regions. Current methods for obtaining these boundaries are principally manual and subjective. This includes hand-drawn ROI analysis where investigators draw polygons using a mouse over one of many potential 2-dimensional MR images. Manual tracing of white matter structures in the brain is very time consuming and requires considerable expertise to accurately identify structure boundaries. For ROI analysis, researchers identify brain regions and compare FA and D_av between research subjects and normal controls (Peled et al., 1998, Kubicki et al., 2002, Mukherjee et al., 2001). Selection bias and variability in the process of selecting images and drawing ROIs introduce significant barriers to both research methodology and clinical assessment.

Investigators have researched and created automated segmentation algorithms of the corpus callosum applied to non-DTI MR images. For example, Lee et al. (2000) proposed an algorithm to automatically find the corpus callosum from a midsagittal MRI using high intensity gray level distributions and a window region growing algorithm as opposed to conventional contour matching. While this algorithm is fully automated, the segmentation sometimes is inaccurate either by not finding the corpus callosum at all, not including part of the corpus callosum, or including other structures in the segmentation such as the fornix. Lundervold et al. (1999) proposed a technique to segment midsagittal sections using multispectral 3D MRI recordings. The algorithm entails using intensity values and a corpus callosum template to segment the corpus callosum from a midsagittal slice. It was demonstrated successfully on 10 brains.

Three-dimensional segmentation algorithms of the corpus callosum using non-DTI MR images have also been proposed. For example, Bueno et al. (2001) describes his semi-automatic segmentation algorithm as an immersion simulation modeled by a 3D region adjacency graph. The algorithm entails using contrasted regions from morphological reconstruction constrained by the 3D region adjacency graph to identify 3D regions. Then, the algorithm locates the contours of the region using a watershed transform. The algorithm is described as satisfactory and suggests that it be used with complementary techniques based on deformable models to optimize operation.

Conventional clinical analysis does not include a standardized quantitative protocol that can be applied to a variety of conditions. The proposed development of a Reproducible Objective Quantification Scheme (ROQS) is essential to analyze large cohorts of subjects reliably, and to provide statistical benchmarks for assessing individual results as pathological. There is a manifest need for a rapid automated or semi-automated method of white matter quantification for both individual and group analysis, but no generally accepted solution to this problem. In this report, we develop and apply a novel approach called ROQS to select a midline sagittal corpus callosum ROI that addresses several of the limitations of previous approaches.

Because ROI analysis is the current gold standard for quantitative MRI analysis, ROQS is designed to be a ROI approach, but using a semi-automatic algorithm to address several shortcomings of manual ROI delineation. Additionally, as noted previously, a major limitation of VBA is the need for spatially normalized data sets. ROQS is designed to be an ROI approach that can operate on both non-normalized and normalized data sets. ROQS exploits the fiber orientation information calculated from the diffusion tensor to define the boundary of anatomically distinct white matter tracts in a semi-automated fashion. After the user chooses a seed pixel within the structure, the algorithm uses the properties of this pixel to determine the appropriate boundary. Once the boundary of the white matter tract is determined, values of FA and D_av from pixels within the boundary are automatically processed for analysis. This makes ROQS extremely fast in comparison to manual ROI delineation. Furthermore, the semi-automated ROQS technique is more reproducible and reliable than hand-drawn ROIs. Reproducibility is demonstrated by comparing multiple raters trying to delineate the boundary of the corpus callosum using manual tracing and ROQS. Reliability is demonstrated not only by the inter- and intra-rater variability, but also by the properties of pixels chosen by ROQS in comparison to hand-drawn tracings.