A combination of artificial intelligence and Lung-RADS can increase scan specificity without reducing sensitivity
It is possible to increase the specificity of low-dose CT (LDCT) lung cancer screening programs without chipping away at sensitivity – all it takes is folding artificial intelligence (AI) in with Lung-RADS.
That capability is based on the development of a lung nodule management system, created by investigators from the University of Saskatchewan. The team, led by Scott Adams, M.D., from the department of medical imaging, created a methodology that upgrades or downgrades a radiologist-determined Lung-RADS categorization by applying the AI-produced risk score.
Not only could this system boost specificity, but it could also save money, the team said. Adams’ team shared its results recently in the Journal of the American College of Radiology.
“Using an AI risk score combined with Lung-RADS at baseline lung cancer screening may result in fewer follow-up investigations and substantial cost savings,” they said.
Specifically, specificity could rise by more than 50 percent, and for each patient who underwent CT lung cancer screening, savings hovered around $72.
For more coverage based on industry expert insights and research, subscribe to the Diagnostic Imaging e-Newsletter here.
The team used an AI model developed by Google in 2019 to create the risk score and compared its performance to how well six radiologists did in assessing 3,197 LCDT baseline scans that were considered to be representative of the general lung-cancer screening population. To ensure they matched the 91-percent sensitivity level achieved by the providers, the team used a 0.27 AI risk-score threshold (based on a 0-to-1 scale).
According to their analysis, the team determined their nodule management strategy upgraded 41 (0.2 percent) category 1 or category 2 classifications from radiologists to category 3. On the flipside, they downgraded 30 percent – 5,750 classifications – from category 3 or higher to a category 2.
For the cost savings calculations, the team used the 2019 Medicare Physician Fee Schedule non-facility National Payment Amount to estimate AI strategy-induced management changes. From their assessment, they determined using this model could result in up to $72-savings for each screened patient.
Additional research for other AI thresholds could also beneficial, especially for Lung-RADS category 4 nodules, they said.
“Systematic manual review of cases deemed high or low risk by the AI algorithm should be another area of further investigation to identify new radiologic signs that may potentially improve radiologists’ accuracy even in the absence of AI,” they said.
Ultimately, additional investigations could lead to AI algorithms being used in a similar way to what has been suggested for screening mammography, they said. The tools could be used to pinpoint normal studies that would not necessitate a radiologist review.