Automatic detection of the intracranial boundary has been attempted using a plethora of image processing techniques. This section discusses a selection of these methods. None of the currently available methods completely solves the problem.
The automatic thresholding techniques investigated usually involve four basic steps. First, a histogram of the MRI voxel intensities is produced. Next, a threshold is selected based on the fact that most of the relatively high intensity voxels in the MR image belong to the brain. The threshold is then applied to the image to produce a binary mask. Finally, morphology and/or expert knowledge is applied to the binary mask to remove non-brain regions. The resulting mask identifies the brain and thus, the intracranial boundary.
Suzuki and Toriwaki use iterative thresholding to distinguish brain tissues from others in axial MR slices [49]. An initial mask of the head is created by thresholding at 20% of the maximum grey value. An initial mask of the brain is created by thresholding at 70% of the maximum grey. These thresholds are then iteratively adjusted based on the geometry of resulting masks (ie. the head mask includes the brain mask). This method is ineffective in the presence of RF inhomogeneity and in slices where the brain is not one homogeneous region closely surrounded by the skull.
Li et al. use knowledge-based thresholding in multimodal MRI data to classify voxels into multiple intensity categories [30]. In each axial slice, they compute the centroid of voxels categorized as brain. Next, four points defining a quadrangle are found at the edge of the brain by tracing left, right, up, and down from the centroid to a transition in tissue categories. All voxels outside the quadrangle that are not categorized as brain tissue are then masked to define the intracranial contour. Obviously, this method works only in slices where the brain constitutes one fairly homogeneous region.
Brummer et al. use histogram analysis and morphology to generate a 3D
brain
mask [7]. Using a model of background noise, they
first automatically generate a mask of the head and perform intensity
correction on the masked volume. Next they create an initial brain
mask by automatically thresholding based on a presupposed brain voxel
intensity distribution. They then eliminate regions in the brain mask
that are too close to the edge of the head. Finally, they use novel
morphological operations to clean up the resulting mask. This method
misses brain tissue in extreme slices and includes non-brain tissues
in others. In some cases it produces errors near the eyes. The method
relies heavily on availability of phantom data for intensity
correction (see Section 3.3.1), so can't be used
retroactively.
Pannizzo et al. detect the intracranial boundary in axial MRI slices by tracing a horizontal line outwards from the centre of the image [36]. The point, in each direction, at which the voxel intensity under the line drops below a reference threshold is considered to be a point in the intracranial boundary. A running average of voxel intensities under the line is then computed. The intracranial boundary points are relocated to the first voxels with an intensity too far below the average. The running average procedure is repeated for rows in the image. The entire process is then repeated for all columns. The result is a sequence of points defining the intracranial contour. This method can only detect the intracranial boundary in slices where the brain is one homogeneous region and may not cope well with RF inhomogeneity.
Lim and Pfefferbaum [31] and Attwood et al. [4] trace lines radially outward from the approximate centre of the head in an approach similar to that proposed by Pannizzo et al. Attwood et al impose a smoothness criterion on the resulting contour to refine the location of the intracranial boundary. Zijdenbos et al. fit a surface to the voxel intensities under the contour produced by [31] in order to produce a local threshold [55][54]. The radial line tracing algorithm is then repeated using the threshold to choose the voxels at the intracranial boundary. Zijdenbos' method is more robust than the others in the presence of RF inhomogeneity and partial volume effects. Clearly, none of these methods works for slices where the brain is separated into two or more disjoint regions.
Cline et al. segment the brain from MR images of the head using statistical classification [11]. To segment the brain, samples of brain voxels and non-brain voxels are interactively identified. Bivariate normal distributions, corresponding to the different tissue types in the PD-weighted and T2-weighted MR images, are fitted to the sampled intensities. All the image voxels are then classified according to where their intensities lie in the distributions. Finally, the results are smoothed to remove discontinuities in classified regions. This method requires user interaction, fails in the presence of RF inhomogeneity, and falsely classifies non-brain regions, such as the eyes, as brain.
Lachmann and Barillot isolate brain tissues in MRI slices using several texture analysis methods [28]. They use texture information to create an initial voxel classification. Cluster analysis and Bayesian relaxation is then used to refine the classification. Lachmann and Barillot do not show results for slices containing the eyes or the mouth. We already established in Chapter 1 that Bayesian relaxation-based techniques confuse these features with brain tissue [21][23][22].
Stringham et al. use a statistical relaxation method that incorporates gradient magnitude as well as voxel intensity for brain segmentation [48]. This segmentation algorithm is robust in the presence of RF inhomogeneity, but confuses tissues, such as the eyes, with brain tissue. Further, user interaction is required to seed the relaxation process.
Vinkin et al. use probabilistic hyperstacks for 2D image segmentation [50]. A hierarchical stack of decreasing resolution images is produced by progressively low pass filtering the original. Pixels in high resolution images (children) are probabilistically linked to voxels in low resolution images (parents) according to an objective criterion. Features of interest are identified in a low resolution image and propagated to all siblings, weighted by the linked probabilities. Thus, a statistical segmentation is produced. Identifying the features of interest and thresholding the final segmentation must be performed manually in this method.
Chakraborty et al. combine statistical segmentation and boundary detection to isolate features in MR images [9]. They first segment the images using a method similar to the ICM algorithm [22]. They then use a parametrically deformable shape model algorithm to find the boundary of interesting features in the segmented image [47]. The shape model algorithm modifies the shape of a pre-defined closed contour to match the shape of a region of interest. This method requires user interaction to seed the segmentation and provide an initial shape contour. Further, the segmentation may fail due to RF inhomogeneity.
Snell et al. use an active surface template to find the intracranial boundary in MRI volumes of the head [46]. The method is based on the active contour model algorithm, ``Snakes'', first proposed by Kass et al. [27]. Given an initial estimate of an object boundary, ``Snakes'' approaches the actual boundary by solving an energy minimization problem. In Snell's method, the user identifies points in MR image that correspond to points on a ``standard'' active surface template of the brain. Based in these points, the template is registered to the image. The ``Snakes'' algorithm is then used to attract the surface template to the intracranial boundary. Snell's method appears to work better than all the other methods discussed herein. However, Snell used high resolution isotropic 3D MRI data to test his algorithm. Such MRI scans are generally not performed clinically. Still, the method requires user interaction and may fail for images that do not contain the entire brain.