Can Artificial Intelligence Improve the Accuracy of Screening Mammography Assessment?
Posted: Saturday, July 11, 2020
A diagnostic accuracy study published in JAMA Network Open found that the addition of deep learning algorithms to radiologist mammography interpretation may result in improved overall accuracy in cancer screening. The study sought to determine whether artificial intelligence could close the gap left by limitations in human mammography interpretation.
“Integrating artificial intelligence to mammography interpretation in single-radiologist settings could yield significant performance improvements, with the potential to reduce health-care system expenditures and address resource scarcity experienced in population-based screening programs,” noted Thomas Schaffter, PhD, of Sage Bionetworks in Seattle, and colleagues.
Between September 2016 and November 2017, more than 1,100 participants forming 126 teams and representing 44 countries, developed deep learning algorithms focused on mammography interpretation. Beginning on November 18, 2016, these algorithms evaluated 144,231 screening mammograms from 85,580 women in the United States and 166,578 screening mammograms from 68,008 women in Sweden. They used images alone, combined images, previous examinations, as well as clinical and demographic risk factor data before producing an overall score that correlated with either a yes or a no for the presence or development of cancer within 12 months.
The top algorithm performed at a lower specificity than that of radiologists in community practice, at 66.2% versus 90.5% (United States) and 81.2% versus 98.5% (Sweden). This algorithm resulted in an AUC of 0.858 for patients in the United States and 0.903 for patients in Sweden; however, the best overall results occurred when the top algorithms were combined with human interpretations performed by radiologists in the United States. These assessments resulted in AUC of 0.942 as well as an improved specificity (92%).
Disclosure: For full disclosures of the study authors, visit jamanetwork.com.
