Concussions are an ever-growing concern to those who suffer head trauma. They’re also a concern to doctors trying to determine if a hit to the head has actually resulted in a concussion. Making a correct diagnosis is critical to keeping the patient out of harm’s way long enough for them to recover.
“If the impact isn’t caught on camera or closely witnessed, we have to rely on the patient’s own report and memory of the incident,” says Yvonne W. Lui, MD, associate professor of radiology and associate chair for artificial intelligence in the Department of Radiology. “But self-reporting by concussion victims is almost by definition incomplete. Sometimes, all they can tell you is that they don’t remember what happened, but their head hurts.”
What physicians need is a more quantitative measurement of physical and metabolic changes in the brain after an impact, explains Dr. Lui. It’s easy enough to scan the brain with a CT or MRI, but physicians don’t really know what to look for in the resulting images. “Sometimes, the scan looks completely normal, even after someone’s head has bounced on the ground hard,” she says. “If it’s a football player, he may end up back in the game in the fourth quarter. But it’s hard to believe nothing has happened to his brain.”
More than 1.7 million Americans are diagnosed with a concussion each year, but Dr. Lui suspects that many of these traumas go undiagnosed, leaving patients at risk for more serious damage.
To provide a more quantitative diagnostic technique, Dr. Lui collaborated with Yao Wang, PhD, of the NYU Tandon School of Engineering, and they’ve developed an AI program that has learned to detect changes in patches of white matter throughout the brain. White matter, which is composed of nerve fibers that connect different parts of the brain and nervous system, stands out in scans and, in many cases, shows signs of alteration after head trauma.
By allowing the software to study the results of some 200 research MRI brain scans of patients who experienced head trauma, along with physicians’ determinations of whether or not a concussion occurred, the software learned to find patterns in the white matter that seem to be associated with concussion.
When the program analyzed the scans without the benefit of a corresponding clinician diagnosis, it arrived at the correct diagnosis 87 percent of the time. “Because of the uncertainties of diagnosing concussion, that’s as high a rate of accuracy as we could expect,” says Dr. Lui.
Beyond the potential clinical benefits of her system, Dr. Lui has applied an ingenious solution to the perennial data shortage that hampers many AI projects. Her project was limited to some 200 scans of suspected concussions. That may sound like a lot, but AI systems often require tens of thousands of cases to build reliable accuracy.
To get the most out of the low number of samples, Dr. Lui took advantage of the fact that concussion damage to white matter is typically scattered throughout different parts of the brain. She divided individual scans into hundreds of pieces, with each patch serving as a unique sample of white matter. In this way, she could feed the system what equated to more than 100,000 samples.
The approach seemed to work well for concussions and may prove a useful strategy for other AI systems in medicine. “A lot of cutting-edge AI research in medicine lacks access to, say, 50,000 samples,” says Dr. Lui. “This approach may be a way to get meaningful results from smaller data sets in some other imaging applications.”
Of course, any AI system is only as good as the data it churns, and Dr. Lui recognizes that her own program must evolve as quickly as the science. “I hope and expect that within five years or so, we’ll have a wholly different way of diagnosing concussion,” says Dr. Lui.
When the field gets to that point, she adds, the next generation of AI systems like hers will be ready to step in and learn from more and better case examples. “I think these systems will be able to recognize not only structural changes in the brain, but also metabolic changes that are harder to detect,” she says. “That could help take the uncertainty out of diagnosing concussions.”