In studies of brain pathology, such as multiple sclerosis (MS) , regions of interest (ROIs) that must be well defined are often examined in detail in magnetic resonance images (MRIs). Traditionally, ROIs are outlined manually by a skilled operator using a mouse or cursor. Computer-assisted methods are used for specific applications such as extraction of MS lesions from MRI brain scans [18,36], or extraction of the cerebral ventricles in schizophrenia studies . In many cases, the computer-assisted tasks need to segment the whole brain from the head. Typically this may be required either because the whole brain is the ROI, such as in studies of alcoholics  or Alzheimer's patients , or because automated extraction using statistical methods is facilitated if the skull and scalp have been removed from the image . In this chapter we present a fully automated method we have developed that is in common use in our research setting . The chapter also includes an overview of several other methods for automated segmentation of the brain in MRI.
Fully automated segmentation algorithms have to set their parameters such as thresholds automatically, they must address the partial volume effect in a reasonable manner, and they must work in the presence of typical radio-frequency (RF) inhomo-geneities. Fully automated methods must be robust — in other words, they must provide consistent segmentation on images acquired from any MR scanner using different fields-of-view, relaxation times, and slice thicknesses.
Hybrid methods that include both image-processing and model-based techniques are particularly effective for brain segmentation [1,4,19,21]. The hybrid method  presented in this chapter starts with a thresholding step followed by a morphological erosion to remove small connections between the brain and surrounding tissue. It removes eyes and other nonbrain structures with a model-based approach followed by more image processing consisting of a morphological dilation to recover some of the eliminated tissue. A final refinement of the brain contour is achieved by an active contour algorithm .
In our method, the threshold for an initial segmentation is computed automatically by applying an anisotropic diffusion filter to the image and using the resulting voxel intensity histogram. The method, which is based partly on 2D data and partly on 3D data, operates best on routine axially displayed multispectral dual-echo proton density (PD) and T2 (spinspin relaxation time) sequences. This method has been successfully used to segment the brain in each slice of many head images from many different MRI scanners (all 1.5 tesla), using several different spin-echo images with different echo times. Examples of brain segmentations obtained with this method are shown in Fig. 1.
This method also works well on axial and coronal 3D T1-weighted SPGR (Spoiled Gradient) sequences. However, on sagittally displayed 3D T1-weighted images, it cannot be used in a fully automated manner because such images require accurate localization of cortical convolutions. In these images parameters have to be adjusted to ensure that the thin dark brain areas will be included and to keep the cerebellum
Copyright © 2000 by Academic Press.
All rights of reproduction in any form reserved.
attached to the rest of the brain, which has to be separated from the back of the neck tissue and the cheeks. Sagittally displayed images can be segmented with other techniques such as those described in [1,10,14,16,19].
relatively higher intensity than other tissue in MR images constitutes the second type of prior information. Using the anisotropic diffusion filter on T2 (or PD) images, the majority of the tissue other than the brain can be darkened, allowing a simple threshold to be used subsequently for segmentation.
Was this article helpful?