PROVIDENCE, R.I. [Brown University] — An artificial intelligence (AI) technique developed by Brown University mathematicians may soon help doctors assess the severity of vascular problems like blood clots and aneurysms.
The technique uses machine learning algorithms that are encoded with the physical laws that govern flows of fluids. The algorithms make it possible to infer key attributes of a flow — velocities, pressures and other parameters — just by analyzing images or short videos of that fluid in motion. In a paper published in the Proceedings of the National Academy of Sciences, an international team of researchers showed that the technique can be used to study blood flow in microaneurysms, tiny bulges in blood vessels that often occur in the eyes of people with diabetes.
“The idea here is to develop a system that can help doctors to determine how severe these aneurysms are and what’s the likelihood of a rupture,” said George Karniadakis, a professor of applied math and engineering at Brown and a study coauthor. “That could help in determining when to intervene before severe damage is done.”
Karniadakis and his colleagues first reported their AI system last year in a paper published in the journal Science. The system makes use of artificial neural networks, sets of computing nodes that roughly mimic the connections made by neurons in the human brain. Neural networks are now commonplace in modern AI, but Karniadakis’ system adds a key modification. Rather than starting from a blank slate, the new system comes “pre-loaded” with the equations that govern fluid flows. As a result, these physics-informed neural networks can quantify the motion they see in just a few images or a few frames of video.
For this new study, Karniadakis worked with researchers from Nanyang Technological University (NTU), Singapore, and Massachusetts Institute of Technology. The team turned the AI system loose on microfluidic devices, microchips etched with tiny channels through which minute amounts of fluid can flow. In this case, the microfluidic devices mimicked tiny blood vessels with microaneurysms of various shapes and sizes. The chips provided training data for the AI system, which enabled the system to then make predictions about pressures and stresses that the blood flow produces on the walls of the device. The research found that the AI algorithm outperformed other methods in determining critical attributes of the flow.
Study co-author Subra Suresh, a former Brown professor who is now president of NTU, said the system could eventually have important clinical applications.
“Currently, measuring the mechanics of blood flow in the smallest blood vessels requires sophisticated equipment and trained personnel,” Suresh said. “With this platform, we can now gain important mechanical information and insights into disease evolution mechanisms that were previously very cumbersome to extract.”
Karniadakis says the team hopes to develop a version of the AI system that could be made available directly to medical professionals. That system doesn’t require a tremendous amount of computing power, enabling it to run on virtually any modern laptop. The idea is to combine the AI system with state-of-the-art imaging technology of blood flow in the eye.
“There are fantastic techniques that produce detailed images of blood flow in real time, but it’s very hard even for an expert to infer pressures and stresses just by looking at them,” Karniadakis said. “This system can do that, and that could be helpful to doctors making treatment decisions.”
Other co-authors of the research were Shengze Cai, He Li, Fuyin Zheng, Fang Kong and Ming Dao. The research was supported by the U.S. Department of Energy’s Physics-Informed Learning Machines Project (DE-SC0019453) and the National Institutes of Health (R01HL154150).