Professors Alan McMillan, PhD, Associate Professor in the UWSMPH Department of Radiology, Po-Ling Loh, PhD, Associate Professor in UW’s Department of Statistics, and Varun Jog, PhD, Assistant Professor in the UW College of Engineering were awarded an R01 from the National Library of Medicine. The title of their project is “Can Machines Be Trusted? Robustification of Deep Learning for Medical Imaging.” The goal of their research is to study ways to strengthen deep learning networks by looking at the ways in which they fail in both intentional and real-world situations. This research will be funded from the National Library of Medicine FOA [PAR18-896] – NLM Research Grants in Biomedical Informatics and Data Science.
Professor Alan McMillan, PhD, received a pilot grant that helped bring the idea of the development of robust AI. “We have always been able to engineer good solutions with AI for any problem… however, it is not known how the methods will respond in respect to unanticipated inputs” Professor Alan McMillan also stated when talking about how they came up with the idea. AI is very specific to the data and application that was used to train the model, so what works well at UW may not work at other institutions for many reasons. Some of the causes of the AI “failures” is that there is data that it hasn’t seen before, or patient effects like motion, operator error of the scanner, or other artifacts could cause an AI algorithm to give an unpredictable result.