PNNL

Can computer vision help radiologists find signs of potential breast cancer?  In an ongoing project, PNNL researchers are discovering to what extent deep learning—a form of artificial intelligence—can be applied to help find cancer in diagnostic images.

In work that began under the National Security Directorate’s Deep Learning for Scientific Discovery Agile Investment, PNNL researchers are collaborating with the Fred Hutchinson Cancer Research Center and the University of Washington to apply data science to human biology. PNNL is training a computer vision model to look for cancer in 10,000 anonymized MRI images of UW patients whose scans were reviewed by radiologists. The goal is to see how well the computer findings match those of expert humans.

Gary Gilliland, FHCRC director, highlighted the PNNL collaboration and future vision in a January 2019 article. “Our hope is to improve early detection, particularly for women with dense breasts,” he said. Dense breast tissue can hide or mask cancers in mammograms. 

Now funded under a PNNL Quickstarter project, PNNL’s initial research appears promising. “In some cases, the deep-learning results are approaching radiologist accuracy,” said principal investigator and data scientist Kayla Duskin. Results are expected later this year.

In addition to Kayla, previous contributors included Court Corley, Nathan Hodas, Shant Mahserejian, and former PNNL staff member Steve Rysavy, all from NSD.