Algorithm hunts for unreported osteoporosis markers on chest x-rays

Thursday, November 30 | 8:50 a.m.-9:00 a.m. | R1-SSCH09-5 | Room E352

An AI algorithm can find radiographic markers for osteoporosis that are common but often not reported on radiology reports, according to this scientific paper.

To date, AI models in medical imaging have often commonly focused on radiological finding detection to provide direct diagnostic assistance to radiologists, according to presenter Jonathan Stephan Luchs, MD. This recent study, however, investigated a use case to enhance osteoporosis risk characterization.

The presence of osteopenia and spine wedge fracture are strong indicators of osteoporosis. Detecting these unreported findings could help characterize risk and improve patient care and management, according to the researchers.

The group used AI to retrospectively process chest images to identify osteopenia of the spine and spine wedge fractures that were not reported in the original radiologist report. The dataset consisted of 519 chest x-rays from patients ages 65 and older, collected from outpatient clinics.

As a comparison, the original radiologist report was manually reviewed for the presence or absence of the findings. In cases where there was a discrepancy between the AI model and the report, a radiologist adjudicator evaluated the CXR scan to determine if the finding was overcalled or undercalled by the AI model/radiologist, Luchs wrote.

The researchers found significant variation between the AI model and the original radiologist report. The AI algorithm reported 76 cases of osteopenia and 58 cases of spine wedge fracture, compared with eight cases of osteopenia and no cases of spine wedge fracture in the original report.

In discrepant studies, comparison with the ground truth - as determined by radiologist adjudicator -- indicated that there was predominantly undercalling in the original report.

"... these findings have ramifications for billing and revenue capture, pre-emptive treatment and preventing patient morbidity and mortality," the authors concluded.

Page 1 of 2
Next Page