Artificial intelligence (AI) and machine learning may concern many in the medical world, but the speakers at the final plenary of SIR 2018 said that the technological advances will help improve patient care and expand care to people around the world lacking access.
A trio of presenters from the Stanford University School of Medicine and Stanford University Medical Center talked about the impact of machine learning and AI on medicine during the plenary on Wednesday, March 21.
The session also included the presentation of one of the abstracts of the year, “Sparing of collagen and extracellular matrix protein in irreversible electroporation-treated normal porcine lung promotes T-cell and macrophage infiltration throughout ablated tissue” by first author Masashi Fujimori, MD, PhD, Memorial Sloan Kettering Cancer Center.
During “Big data, Watson, and why your diagnostic colleagues may not have jobs in 10 years,“ Matthew P. Lungren, MD, provided an overview of “machine vision” applications in diagnostic imaging, the challenges in those applications and the machine vision research taking place at Stanford.
AI has become excellent at classifying images, said Dr. Lungren, assistant professor of radiology (pediatric radiology) at the Stanford University Medical Center. Every year, the error rate decreases and now it is better than humans at classifying.
Stanford started the Artificial Intelligence in Medicine and Imaging (AIMI) laboratory to develop and evaluate AI systems that benefit patients. Positive results have been seen so far in detecting fractures, congenital abnormalities in pediatric neurology and deep-vein thrombosis.
In one test, 12 expert radiologists from around the nation examined results to see how the computer model compared with the human’s work. It did just as well as the humans, but the human radiologists took three-and-a-half to four hours to complete their work. The computer model took one minute.
But medical imaging files are large and complicated, and machine learning still struggles—especially in areas involving finding the context in the images.
“It is really difficult for software engineers, machine engineers to understand the concept that there is gray area,” Dr. Lungren said. “There’s no black and white in a lot of diagnostics.”
How will this affect the workforce? Dr. Lungren said it’s possible that efficiency could increase so much that some jobs would be eliminated.
“But here’s my favorite scenario,” he said. “We’re working on taking this AI — and there’s going to be a lot of market forces at play — but we’re planning on releasing every model we build for free. A lot of people don’t have access to proper health care, and that’s why these AI tools are so exciting. If you can provide diagnostic services to people who don’t have access, you can do a lot of good.”
Justin Ko, MD, MBA, medical director and chief of medical dermatology for Stanford Health Care, discussed the application of deep learning in dermatology. Stanford has applied the technique to both general skin conditions and specific cancers.
Early tests comparing AI with dermatologists screening skin lesions found that AI was at least as accurate, and in most cases, more accurate in diagnosis. His group has started using AI in a clinical setting to validate the previous results.
Dr. Ko sees AI as a way to augment, not replace, the clinician.
“I think it will really allow us to leverage the expertise of the individual clinician, increase productivity, increase access, increase the ability of our field to take care of skin issues and do it in a way that’s more accessible and actually better for the patients,” he said.
Session co-coordinator Lawrence V. Hofmann, MD, FSIR, talked about his experience with the company he co-founded, Grand Rounds, which collects data and uses machine learning to help patients find the right physician using factors such as distance from patient, quality and insurance.
Dr. Hofman pointed out a study that found that almost half the patients who had Whipple procedures had them done by a surgeon who does on average less than five of the procedures a year.
“That’s crazy. We know that surgical volume is one of the best predictors for outcomes,” he said. “But what’s even more concerning is when you actually look at how far these patients live from someone (with more procedure experience), more than 90 percent of those patients could have gone to a doctor closer who had more experience.”