Functional MRI and AIR

For their MRI study, Cohen and his colleagues used a conventional MRI machine, like those that became available in many hospitals during the 80s, a big advantage since the research can be accomplished without major new investment in technology. Functional MRI works on the principle that when brain cells (neurons) become active, blood flows to them, and the MRI scanner registers increased oxygen in the area. Because MRI machines used in this way detect changes resulting from biological function, the method got its name.

The technique generates large amounts of data quickly -- a great advantage, says Cohen, and a problem. "It gives a lot of information to work with, but likewise it's a tremendous amount of data to process -- as much as half a gigabyte per experiment." To deal with the data overload, Cohen and his colleagues last fall turned to the Pittsburgh Supercomputing Center's Alpha Cluster, a linked network of 14 DEC Alpha workstations. They used the cluster to address a particular problem of their functional MRI experiments. A human subject stays in the machine for two to three hours as the MRI scanner records data. Though special pillows are used to reduce movement, it's impossible to keep the head perfectly still. Software called automatic image registration (AIR), developed by Roger Wood of UCLA, can correct for head movement, but the sheer number of images -- typically 1,200 per experiment -- creates an imposing demand on computing.

AIR is an ideal application for the Alpha Cluster, notes Cohen, because it is inherently parallel. A single experiment typically records 200 separate images for each of 6 separate scan sites, or slices. "For each slice," says Cohen, "you take as a reference point one of the 200 images and align all the others to it. There's no need for communication back and forth. Each sample can be aligned to the reference independently."

On a high-speed workstation, says Cohen, it takes as much as 24 hours computing time to register the images from one experiment. "Often, we run two experiments in an evening, which means the computing can't keep up with the data, and this cripples the research." On the Alpha Cluster, the same computing takes an hour, a radical speedup that overcomes the research bottleneck. Having successfully adapted AIR to a clustered workstation environment, Cohen and his colleagues hope to make the software available as a community resource for functional MRI research.

Looking Ahead

A series of computational steps are involved in turning MRI data into a useful picture of the brain at work. Prior to the image-registration processing done by AIR, the raw data from the MRI scanner, which is recorded as a varying intensity of radio waves, must be converted into a physical image of the brain. This step also requires significant number-crunching, notes Cohen, and the researchers are looking into moving this processing also to the Alpha Cluster so the MRI data can flow directly into AIR.

After AIR, further processing involves statistical analysis and volume rendering of the images. Looking ahead, Cohen and Walter Schneider, of the University of Pittsburgh's Learning Research and Development Center, envision tying their functional MRI experiments directly to the Pittsburgh Supercomputing Center. With sufficient computing power, images of the functioning brain could be available in real time as a clinical tool in diagnosing and treating disturbances in brain function and cognition.

go back to the main screen