Background

Global quantification of tissue volumes from MR brain images is typically accomplished using segmentation methods that

Copyright © 2000 by Academic Press.

All rights of reproduction in any form reserved.

attempt to classify each voxel in the image as GM, WM, or CSF. This classification is based predominantly on intensity information from the image. In MR images, contrast between tissue classes can vary according to the pulse sequence of the acquisition. Furthermore, multiple MR images with varying contrasts can be acquired of the same subject, thereby yielding additional information for separating classes. The type of MR data, as well as artifacts that may be present, can have significant effects on segmentation performance.

Much of the literature has focused on quantifying volumes using multi-spectral data like double-echo images formed by spin density and T2-weighted sequences [1-7] or highresolution Tl-weighted scalar data [8-12]. Figures 1a and 1b illustrate a typical double-echo imaging of the brain, whereas Fig. lc depicts a Tl-weighted image of the same brain. Figure 1d shows an example of a segmentation computed using the Tl-weighted image.

Double-echo images have been widely used in clinical settings. They provide excellent contrast between CSF and brain tissue. However, these images generally have thicker slices and lower GM/WM contrast than high-resolution scalar acquisitions. Tl-weighted images, on the other hand, are becoming the de facto standard for MR volumetrics. They provide high-resolution data, good tissue contrast, and low noise without major increases in acquisition time. One of the main disadvantages of these images, however, is that contrast between CSF and nonbrain tissue (e.g., dura) can be poor, presenting difficulties in separating these tissue classes [l2].

Two artifacts that can significantly affect the performance of segmentation methods are partial volume effects and intensity inhomogeneities. Partial volume effects occur where multiple tissues contribute to a single voxel in the image, resulting in a blurring of tissue boundaries. These artifacts can be modeled using soft segmentation methods, which allow tissue classes to overlap, rather than exclusively assigning a tissue class to a voxel. There have been few results, however, demonstrating the application of soft segmentations for volumetric quantification purposes [l3, 34]. Instead, soft segmentations typically are converted into hard segmentations before volumes are measured. Intensity inhomogeneities cause a shading artifact to appear over the image, degrading the performance of methods that assume the intensity of a tissue class is constant over the range of the image. Attempts to compensate for these artifacts use either a prefilter that removes the inhomogeneities [3, l4-l6], or an adaptive segmentation that simultaneously compensates for the shading [8,9, l2, l7, l8].

In the next section, we describe several segmentation methods. Although each method is described separately, different techniques may be used in conjunction with one another to optimize the segmentation of a scene. In addition, removal of unwanted regions present in the scene may considerably improve the speed and accuracy of the segmentation. In particular for MR images of the brain, removal of the skull (bone), muscles and fat greatly increases the algorithms'

efficiency and improves the segmentation accuracy of brain tissue and CSF. In this case, the elimination of an unwanted tissue type (e.g., bone) may be accomplished by a sequential application of morphological operators, thresholding and seeded region growing [l2].

0 0

Post a comment