|
||||||||
|
RADIOLOGY Computer Method Speeds Labeling Of Brain StructuresDetects Disease Changes In a darkened room, Megan Dieterich sits at a computer workstation and looks at black and white cross-sections of a human brain all day. Like the other technicians who work at the Center for Morphometric Analysis at Massachusetts General Hospital, she is patiently divvying up the fuzzy masses and folds of tissue from the MRI scans into the precisely delineated regions of scientific convention--pallidum, putamen, thalamus, ventricles. "It takes a long time," Dieterich said, which is true: it will take her more than a week to chart the regions of one brain, even with the help of software that can draw some of the borders automatically. New employees at the center must complete a daunting three months of training before they can even begin working, and demand for their services is high enough that a stack of brains is waiting for them when they begin. "Your first brain can take a month," she adds.
The brain-labeling technique devised by Anders Dale, Bruce Fischl (above), and colleagues illuminated structural changes associated with Alzheimer's disease. Photo by Graham Ramsay An accurate segmentation of the brain in MRIs is needed to study differences in the size and morphology of brain structures in diseases like Alzheimer's or schizophrenia, as well as functional imaging studies that tie brain activity to specific structures and locations. Developing a faster, automated method of labeling the structures in the brain has been a much desired but elusive goal. In the Jan. 31 Neuron, a team led by Anders Dale and Bruce Fischl, both HMS assistant professors of radiology at MGH, demonstrate a computer technique they have developed that can label all the structures in the human brain rapidly and with an accuracy comparable to manual segmentation. As a first test, they were able to detect subtle changes in volume and morphology associated with Alzheimer's disease. Teaching the ToolThere are some tasks at which computers naturally excel, but translating Megan Dieterich's subtle decision-making into code is difficult. Much of what is gleaned from the MRI images is based on brightness--the differing intensity of white matter and gray matter, the darkness of a fluid-filled ventricle, or the outline of the folds of the cortex. But many structures are not conveniently marked with discrete intensities, and drawing borders between them requires more finesse. "Just knowing how bright something is is not sufficient to allow you to say whether it's amygdala or it's hippocampus, so we had to encode other types of information," said Fischl.The MRIs are divided into millions of voxels--3-D units akin to pixels in a 2-D image--and each of these can be assigned to a structure based on a variety of factors. By gleaning information from a training set of manually labeled images, a computer can use probabilistic information to assign values to different voxels. One of the factors often used to categorize structures is the global position of a point in the brain. However, this approach relies on aligning brains in a similar configuration, which is problematic because brains vary in size and shape. "The other piece of information we used that really makes the problem tractable is that the structures occur in this very stereotypical spatial relationship to one another," Fischl said. After all, the exact position of the hippocampus within the image may shift, but it will always be positioned behind and below the amygdala, never in front of or above it. A trick widely used in image processing to take advantage of spatial relationships in calculating probabilities is Markov random fields. According to this approach, the probability of a value occurring at a given point depends not only on properties of that point but on the properties of its neighbors. The value of a pixel in an image, for instance, can be deduced in part by the values of surrounding pixels. Normally the approach weighs all these probabilities equally, but in order to make use of the relationships between brain structures, the team varied these probabilities based on direction--the probability that a voxel below the amygdala is hippocampus is much higher than one above it. These constraints narrow the choices for each voxel considerably. "You can iteratively recompute all the probabilities and say what's the most likely label, and then you wind up with a very accurate segmentation," said Fischl. Computer DiagnosisOf course, one of the dangers of relying on probabilities is the tendency to discount what is improbable or out of the ordinary. In order to label accurately a set of brains of patients with Alzheimer's disease, the team modified the training set of images to adjust the norm from which the probabilities were taken. The researchers were able to detect subtle changes in the brains compared to controls.While several studies have been able to measure overall volume changes in the brain due to disease, this technique demonstrates the possibilities of detecting localized structural differences in more extensive studies. Marilyn Albert, HMS professor of psychology in the Department of Psychiatry at MGH, who provided the team with the patient data and scans, is interested in studying these changes in Alzheimer's disease and ultimately using imaging as a tool to track and predict the course of disease and as a better way to determine how well medications are working. "What we're hoping to do is get much more data," she said, because individual variations make it impossible to generalize from just a few examples. In addition, Anders Dale hopes the technique can find its way into the clinical setting, "thus providing a complement to the largely qualitative and subjective methods currently used in evaluating neuroradiological images." His group has been collaborating with several teams to explore computer-aided diagnosis in neurological diseases like Alzheimer's, schizophrenia, Huntington's, and multiple sclerosis. --Courtney Humphries |
