AI helps radiologists detect more breast cancers
Artificial intelligence (AI) support systems for reading mammograms can improve a radiologist’s ability to detect cancers, according to a new study published in Radiology. Using the system, the authors added, does not lengthen the overall reading time.
The study included screening mammograms from 240 women performed between 2013 and 2017. The median patient age was 62 years old. The 240 examinations included 100 that showed cancers, 40 that led to false-positive recalls and 100 from healthy patients.
The AI support system used in the study was designed to automate breast cancer detection in both mammography and digital breast tomosynthesis. It can be used with equipment from numerous vendors and was “trained, validated and tested” with more than 9,000 mammograms with cancer and more than 9,000 without cancer, according to the study. During the study, a team of 14 radiologists interpreted the exams twice, once without AI support and once with AI support.
“During each session, radiologists read half the examinations with AI support and half unaided,” wrote Alejandro Rodríguez-Ruiz, of Radboud University Medical Center in the Netherlands, and colleagues. “Radiologists were blinded to any information about the patient, including previous radiology and histopathology reports. Before the first session, each radiologist was individually trained in a session with 45 examinations not included in the final evaluation. The training was intended to familiarize radiologists with the evaluation workstation, the evaluation criteria, and the AI support system.”
Overall, the area under the receiver operating characteristic curve (AUC) for the radiologists was 0.89 with AI support and 0.87 without it. Sensitivity increased from 83 percent to 86 percent with the AI support, and specificity increased from 77 percent to 79 percent.
Rodríguez-Ruiz and colleagues added that using AI helps radiologists detect more cancers while also improving a radiology department’s efficiency.
“Given the high workload of screening programs, from a cost-effectiveness point of view the performance benefit of using AI support is further enhanced by the fact that radiologists do not lengthen their reading time when using this system,” the authors wrote. “In fact, in a real screening scenario, the average reading time per case would actually decrease by approximately 4.5 percent.”
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.