Researchers from the University of California, San Francisco (UCSF) set out to determine the feasibility of creating a large annotated dataset for machine learning using only routine clinical data. Making use of various natural language processing and machine-learning techniques, they were able to ascribe ground-truth radiology and pathology results to more than 700,000 mammograms, according to presenter Dr. Hari Trivedi.
"Using this dataset, we have been developing algorithms for the detection of breast cancer and also identification of negative studies with high confidence," Trivedi said. "Breast imaging is a great place to start with machine learning in radiology due to the inherent structured data and large volume of exams done annually."
In testing, the researchers found that their current algorithm is able to identify negative studies with a negative predictive value of 99.7%.
"This could serve as a second reader alongside radiologists or as a standalone reader automatically clearing negative cases, thereby allowing the radiologist to dedicate time and effort toward more complicated studies," Trivedi told AuntMinnie.com.
Head over to Lakeside Center at McCormick Place on Wednesday to take in this presentation.
