Thyroid cancer is common, but many cases — up to 77% — are believed to be over diagnosed. That means that people with cancer that may never have caused harm or symptoms may undergo stressful and costly unnecessary biopsies.
Researchers in the Duke University Department of Biostatistics and Bioinformatics and the Department of Radiology aim to solve this problem. They will develop an artificial intelligence (AI) model that will distinguish high-risk thyroid nodules from low-risk ones, thanks to a new grant from the National Institutes of Health.
“This is a paradigm shift: rather than focusing on whether nodules are benign or cancerous, the tool will determine whether or not they pose a risk to the patient’s life,” said Maciej Mazurowski, PhD, associate professor of biostatistics and bioinformatics and one of the principal investigators of the grant. He also directs the Duke Spark Initiative, which aims to advance research and implementation of AI in medical imaging.
Mazurowski is collaborating with Benjamin Wildman-Tobriner, MD, associate professor of radiology.
Over four years, the researchers will create a new deep-learning model that will be trained on patient ultrasound images and data, in collaboration with multiple institutions, including Stanford, University of Pennsylvania, UCSF, and UC Davis.
The hope is that the tool will provide clinical guidance that is more accurate and more reproducible than current evaluations, which rely on subjective evaluation of ultrasound images, Majurowski said.