Artificial intelligence, real-life results
CSE and Medical School team is using machine learning to diagnose COVID-19
As the United States still struggles to achieve adequate testing for the novel coronavirus, researchers are finding that medical imaging could provide an alternative and easily accessible means of diagnosis. A University of Minnesota team is leveraging artificial intelligence to speed up that process.
College of Science and Engineering assistant professor Ju Sun has partnered with Christopher Tignanelli, an assistant professor in the Medical School’s Department of Surgery, and Erich Kummerfeld, a research assistant professor at the Institute of Health Informatics (IHI), to develop artificial intelligence technology that can diagnose COVID-19 via X-ray imaging. Other key contributors include Genevieve Melton-Meaux, a professor of surgery at the U of M Medical School and chief analytics and care innovation officer for M Health Fairview; and Tadashi Allen, an assistant professor of radiology at the medical school.
The researchers recently deployed this tech across all 12 M Health Fairview hospitals, allowing quick and accurate screening of COVID-19 patients.
Sun, a faculty member in the Department of Computer Science and Engineering, researches computer vision and machine learning, which consists of training artificial intelligence (AI) to understand and interpret visual information like images. He started collaborating with IHI and the medical school last fall to develop an AI system that could detect rib fractures in chest X-rays. As the coronavirus pandemic hit in March, their team wondered if the research could apply to COVID-19.
“As COVID arrived, we quickly realized that these two problems are so similar,” Sun explained. “The input is still chest X-rays, it’s just that the output is different. One is to detect a fracture, the other is to decide if there is COVID present or not. We realized that COVID is more urgent, and in the short term, it’s the most important problem to solve.”
So, they changed course—and got to work.
Sun’s team, consisting of Ph.D. students Le Peng and Taihui Li and graduate student Dyah Adila, received de-identified X-ray images from M Health Fairview: 18,000 from patients with COVID-19 and 100,000 from patients without the disease. They fed the X-rays into the artificial intelligence system Sun designed, training it discover COVID-related patterns in the images.
Now that the AI is fine-tuned enough to accurately identify COVID-19, the technology has been implemented in M Health Fairview hospitals with the help of Epic Systems Corp., a healthcare software company that provides computing and health data management infrastructure for hospitals across the nation. It is now available for more than 450 health care systems worldwide through an Epic app called Orchard.
In addition to helping radiologists interpret X-ray images, this artificial intelligence system could get patients their coronavirus test results faster.
Since its inception, COVID-19 testing across the United States has been slow-going. But, the turnaround time is due to a health organization’s capacity for testing, not the test itself, since many samples must be carted to external research facilities for processing. Using an AI system would circumvent that trip entirely.
“X-ray is a very accessible process,” Sun said. “It only takes several minutes to get the images."
"The analysis will be quickly run in their analysis system, and you’ll get the result in a matter of seconds,” he added.
However, there are limitations to using artificial intelligence, Sun said. AIs must be “trained in” with a diverse enough data set, otherwise they won’t be able to accurately interpret or predict information. Imagine if you only trained a self-driving car on smooth, paved roads in the summer, then the system would not be able to drive on icy roads during the winter.
With COVID-19, the X-rays could look different across different demographics and depending on how severe the disease is. So, the researchers must continually tweak and improve the AI to detect patterns despite these differences.
Sun believes that once optimized, the AI system will be as accurate as—or more accurate than—the current nasal swab tests.
Now that the researchers have implemented this system in M Health Fairview hospitals, they plan to take it one step further: to use the AI to predict the progression of the disease in different patients.
“Especially for COVID, doctors are interested to know whether the patient will have adverse outcomes,” Sun said.
“This helps to prioritize treatment and optimize planning and allocation of medical resources," he explained. "Maybe after several days, the situation will escalate, and they will need to put the patient in the ICU or consider mechanical ventilation.”
In addition, the University of Minnesota team is working to transfer the model to researchers from Indiana University and Ohio State University so they too can run AI X-ray tests in their respective cities.
While privacy regulations prevent hospitals from sharing their data easily, the researchers hope they will be able to optimize their COVID-19 AI systems without any direct data sharing, based on a new technique called “federated learning.”
“Our UMN-IU-OSU partnership is also going to address a long-term issue of data sharing in the medical world,” Sun said. “This is mostly like a pilot program, so that us three universities can easily train AI models—whether for COVID or for other diseases— making the best use of all available data across the sites, absent privacy concerns.”
Story by Olivia Hultgren
If you'd like to support this work and other COVID-19-related activities, visit the CSE Response Fund.