Algorithm Slashes Compute Time for Low Radiation Dose Scans from 100 Hours per Image to Less Than an Hour
GE’s “Veo” scanner technology helps lower patients’ radiation exposure. Courtesy of GE Health (Flickr photo)
One of the tools doctors rely on to diagnose conditions such as cancer, kidney tumors and spinal trauma is an imposing machine called a computed tomography scanner, or CT scanner.
In a hospital emergency room, standard CT scanners can quickly look over the affected area, and in less than 5 minutes generate images of the inside of a patient’s body, helping doctors make life-saving decisions.
But there’s a catch, says Dr. Ella Kazerooni, a professor of radiology at the University of Michigan Medical School: “The radiation dose for a standard chest CT scan is equal to about 70 chest X-rays.”
That may be OK when it’s a desperate life-or-death emergency, but what about when doctors need to take regular CT scans of a small child with a long-term disease such as lymphoma, or an adult with a brain tumor? The amounts of total X-ray radiation such patients may be exposed to could quickly reach levels that could elevate their risk of cancer.
In those cases, doctors can use CT scans with very low doses of X-ray radiation that dramatically reduce the patient’s cancer risk.
But when they do that, there’s yet another catch: It takes lots of time, and huge amounts of computing power, to turn the smaller dataset from a low-dose scan into a usable medical image. We’re talking not hours but four to five days of computing time on mainframe-equivalent computers to come up with a workable image. For many doctors and hospitals, both the computer power needed and the long delay to get an image have made low-dose scans impractical.
One leading CT scanner vendor, General Electric, was determined to crack that challenge. “It was one of the grand challenges of medicine: How could we crack this problem to yield better images at dramatically lower X-ray power settings, and in less time?” said David Baker, an Intel engineer who worked on the problem.
Cracking the Code
The answer to the question Baker posed lay in a set of mathematical rules called an algorithm.
A human bran viewed with GE “Veo” scanner technology. Courtesy of GE Health (Flickr photo)
Fine-Tuning Software to Save Lives
“It was a multi-year effort,” Sych recalled. “There was a lot of fine tuning. It got down to counting individual clock cycles for each step of the algorithm.” It was difficult, for example, to change the algorithm, which worked best in a single-threaded environment, into one that could take advantage of multi-core processors.
It also helped that while this effort was going on, Intel’s “tick-tock” processor strategy went through two generations. By 2010, Intel had a whole new generation of Xeon processors, which were still socket- compatible with previous Xeon processors.
“That was important,” Johnson said. “GE wanted a solution that they knew they could count on for at least 7 years.”
As Baker described the breakthrough, “The joint team ultimately developed an accelerator based on 28 Xeon processors totaling 112 cores and a dramatically improved algorithm. We reduced the compute time to around an hour, delivering superior medical images and reducing the X-ray power by up to 90 percent.”
Dramatic Reduction in X-Ray Exposure
“We have been able to reduce X-ray doses to previously unthinkable levels,” said Professor Johan de Mey, head of the radiology department of University Hospital in Brussels, Belgium. That is opening up the benefits of CT scans to a wider variety of patients.
Evgeny Drapkin, a principal engineer at GE Health, said that exposure levels for scans done with the Veo machines have been reduced by 4x.
Baker declared, “Kids who have long-term diseases, who have to get regular CT scans to check the progress of the disease, can often approach their lifetime limit of exposure to X-rays. With the Veo scanner they can dial down the X-ray power and dramatically extend the number of times they can get a scan.”