A recent study led by the UW demonstrated that artificial intelligence (AI) could be used to increase the accuracy of single-radiologist assessments of mammograms.
The study showed that overdiagnosis, a major harm of screening mammography, can be reduced by AI, improving quality of care in diagnosing breast cancer by preventing up to 500,000 unnecessary workups. However, clinical testing would be necessary before an algorithm could be broadly adopted, Christoph Lee, professor of radiology at the UW School of Medicine (UWSOM), said in an interview with UW Medicine.
The UWSOM, in conjunction with several other research groups, hosted the Digital Mammography Dialogue for Reverse Engineering Assessment and Methods (DREAM) challenge, which sought to utilize growing AI capabilities to improve the accuracy of diagnosing breast cancer. The event brought together two distinct backgrounds: clinical mammography and data science.
The event utilized two different datasets of mammogram images, one from Kaiser Permanente Washington and another from Sweden’s Karolinska Institutet (KI), for different stages of the event. The first dataset was used during the competitive and community phase of the event and the second during algorithm validation. By using a model-to-data approach where participants submitted models for remote training instead of downloading the data, the event protected the privacy of those whose past mammograms would help future diagnostics.
“It was an iterative process where challengers had access to a very limited amount of testing data and to train their model,” Leesaid. “Then they would send their algorithms to Sage Foundation and it would run on the entire data set and results would be sent back to them.”
The artificial learning model performed differently on the two datasets.
“[The KI dataset] functioned mostly as a validation data set, to see how challenger’s algorithms performed in a European population,” Lee said.
The percent detected differed slightly in the two data sets, which the study suggested was due to differences in the datasets, such as intervals and procedures for screening, not the AI’s performance.
Although the AI performed worse than a human reading of a mammogram, the study showed that AI could be used as a kind of second radiologist. Since Sweden already uses a two-view consensus involving two radiologists to diagnose breast cancer, the AI does not meaningfully improve their diagnostic capabilities. But in the United States, where only one radiologist examines a mammogram, the AI aid improves the chance of a correct diagnosis.
“The most difficult part for me was communicating [with] and finding common ground between the data science community and the clinical radiology community,” Lee said.
The study showed that the process for diagnosing breast cancer could be improved dramatically, easing costs and eliminating unnecessary workups, by bringing together the data science and clinical radiology communities.
“Therefore, if we’re really going to make an impact in the field like cancer screening with AI, it’s going to have to be a very deep partnership where there’s a lot more trust and a lot of communication,” Lee said.
Reach contributing writer Leonard Shin at firstname.lastname@example.org. Twitter: @WombatLeonardS
Like what you’re reading? Support high-quality student journalism by donating here.