Tuesday, December 3 | 1:30 p.m.-1:40 p.m. | T6-SSCH06-1 | Room 451A
Even as the number of U.S.-cleared AI algorithms continues to increase, there's still a lack of data on how they perform in real-world clinical settings. In this session, an international postmarketing surveillance study will be presented that evaluated a commercial chest x-ray AI algorithm's performance in real-world clinical practice.
Presenter Parisa Kaviani, MD, of Massachusetts General Hospital in Boston and colleagues included 2,096 chest x-rays from four sites in India, two sites in Korea, and one site in Thailand where the algorithm (Insight CXR, Lunit) was used alone as well as to aid in interpretation. All images were deidentified and uploaded to an offline, secure annotation platform, where two thoracic radiologists reviewed the images for the following findings: pulmonary nodule, atelectasis, cardiomegaly, consolidation, fibrosis, mediastinal widening, nodule, pleural effusion, pneumothorax, and active tuberculosis.
The overall performance of AI-aided radiology reports was higher than the standalone AI algorithm, notably with small but significant differences in suboptimal images, for instance, incomplete anatomic coverage of the thorax, under- or overexposure, marked patient rotation, overlying external object, low lung volume, and artifacts impairing evaluation.
This study is an RSNA award winner in the Trainee Research Prize – Fellow category. Check out the session to get all the details.
