White Matter Architecture Core

Introduction

The main goal of the White Matter Architecture core is to develop, implement, and apply novel methods for the analysis of Diffusion MRI data (dMRI). dMRI is an MR modality that has gained tremendous popularity the past five years and is especially promising for imaging the white matter in the brain. The extension of available technology will aid in the diagnosis and subsequent treatment of disorders of the central nervous system and is likely to have a major impact on assessment of white matter pathologies (e.g., schizophrenia, multiple sclerosis), quantification of abnormal white matter development, detection of stroke and trauma including traumatic brain swelling, diffuse axonal injury, and spinal trauma, as well as a large variety of brain tumors. In addition to direct clinical impact, the extension of available dMRI technology will also contribute to basic neurosciences, improving our understanding of physiological white matter development, mature brain connectivity, and aging. Several clinical applications will drive the technology development in this core. However, to keep a clear focus, we will continue to use schizophrenia as our main target application. The use of dMRI for finding white matter abnormalities in schizophrenia will likely have a large impact on further understanding of the etiology of schizophrenia, a devastating disorder affecting close to 1% of the population. The work in this core is performed together with our long-standing collaborator Dr. Martha Shenton, a world-renowned expert in the field of MRI findings in schizophrenia.

Diffusion MRI in 3D Slicer

Tract display in Slicer3.

The 3D Slicer (or simply Slicer) software was initially developed as a joint effort between the Surgical Planning Laboratory at Brigham and Women's Hospital and at the Computer Science and AI Lab (CSAIL) Medical Vision Group. The program has evolved into a national platform supported by a variety of federal funding sources. This versatile research environment has resulted in a wide array of functionality, supporting a variety of medical imaging projects.

Slicer is a "point and click" end-user application. Slicer is used as a vehicle for delivering algorithms to computer scientists, biomedical researchers and clinical investigators. Slicer is distributed under an open source license without a reciprocity requirement and without restrictions on use. For a sampling of the portfolio of applications, please see the Slicer Community page.

DTMRI is a module in the 3D Slicer for interactive visualization of diffusion tensor MRI (DT-MRI, or DTI). Its developers are in the Laboratory for Mathematics in Imaging (LMI) at Brigham and Women's Hospital, and the MIT Computer Science and AI Lab (CSAIL) Medical Vision Group.

Slicer3 is a cross-platform end user application for analyzing and visualizing medical images and is a collection of Open Source libraries for developing and deploying new image computing technologies. This platform is intended to replace the venerable Slicer2. This is an algorithm development platform with a powerful new Execution Model to facilitate creation of new modules.

Uncertainty in DT-MRI Tractography

Probability density functions of the underlying fiber orientation for the three different voxels indicated in the FA map to the right.

Conventional tractography methods estimate fibers by tracing the direction of maximum water diffusivity. A main limitation of this traditional approach is that it gives an impression of being very precise. However, in practice there are several factors that introduce uncertainty in the tracking procedure. Noise, splitting and crossing fibers, head motion and image artifacts are all examples of factors that cause variability in the estimated fibers. To address this uncertainty we have been working on stochastic tractography methods, which aim to quantify and visualize the uncertainty associated with the estimated fibers. The figure shows, a probability density functions of the underlying fiber orientation for the three different voxels indicated in the FA map to the right.

DT-MRI Visualization

Local water diffusion, visualized using 3D glyphs, in a zoomed region of interest.(Left) Diffusion estimated using a single tensor model. (Right) Diffusion estimated in the same region using the proposed two-tensor model.

Another line of work is to extend the widely used tensor model of the local water diffusion profile. Because of the relatively large voxel size in DT-MRI, a voxel may contain two or more fiber bundles with different orientations. The widely used tensor model, which assumes a single bundle in each voxel, is unable to describe such a situation. A model mismatch introduces extra variability in any parameter used for quantifying the anisotropy of the water diffusion, which ultimately leads to a loss of sensitivity when testing hypotheses regarding differences between schizophrenics and normal controls. We have developed a two-tensor model (see initial implementation in Peled et al., 2005), which recently was presented at the ISMRM 2005 conference. The traditional DT-MRI single tensor model is shown to the left, and the extended two-tensor model shown to the right. The two-tensor model allows the modeling of crossing fiber tracts.

Grouping DT-MRI Tractography Results to White Matter Bundles

A clustering algorithm takes a number of traced fibers (left), extracts features from these fibers (middle), and produces a segmentation based on the similarity of the fibers (right).

Our initial goal was to color fibers in order to enhance visualization and human perception of fiber trace connectivity. In (Brun et al., 2003) we developed a novel method for pseudo-coloring fiber traces, using a recent machine learning technique called Laplacian Eigenmaps, where fiber traces with similar connectivity were assigned similar pseudo-colors. This method is useful mainly in explorative studies.

Example of the fiber tractography using color-coding. Sagittal view of the white matter fiber tracts demonstrates anatomical connectivity by fibers of similar color.

In the result to the left, showing sagittal and axial views of the white matter fiber tracts demonstrate anatomical connectivity by fibers of similar colors, traces connecting similar areas on the cortical surface were considered more similar than traces connecting different cortical regions. Each fiber is mapped to a 3-dimensional space corresponding to the color components red, green and blue. Hence, color is used here to explain an abstract mapping of each fiber trace onto a three dimensional space using the criteria of similarity.

Preliminary results from population clustering. Corpus callosum (left), Cingulum bundles(middle), and Uncinate Fasciculus (right). These clusters were obtained by a simultaneous clustering of 5 tractography data sets.

In more recent work (Brun et al., 2004), we introduced a method that aims to segment fibers into bundles. This method is based on a graph partitioning method called Normalized Cuts. The proposed method recursively divides clusters into two parts until a satisfactory segmentation has been obtained. A segmentation using this method is shown to the right.

This technique allows us to use fibers from multiple brains as input, and thereby obtain a simultaneous clustering and matching of the bundles in all brains. In addition, we automatically obtain correspondence of bundles across brains; by selecting one or several paths of interest in one brain, the most similar paths in all brains are obtained as the nearest points in the high- dimensional space.

Retrieved from "http://nac.spl.harvard.edu/pages/Research_Cores/WhiteMatterArchitecture"