Sunday, November 29 | 1:00 p.m.-2:00 p.m. | S4-SSCH02-5 | Room E451A
Researchers will describe the performance of a new deep-learning algorithm for predicting malignancy on low-dose CT (LDCT) lung cancer screening exams.
The Brock model is used clinically to help physicians choose the most appropriate management for suspicious lung nodules, but deep-learning algorithms based on imaging alone may offer better risk prediction, according to presenter Maurits Engbersen, PhD, of AI firm DeepHealth, and colleagues. The group evaluated the Kiraly algorithm in an independent screening population and compared its performance with the Brock model and Sybil, another deep-learning model.
The researchers retrospectively gathered 1,191 low-dose CT scans acquired at four U.S. imaging centers between 2016 and 2018. Of the participants, 195 had lung cancer. Exams reported as Lung-RADS 3 or higher were tracked until a cancer diagnosis or two years of follow-up without one, according to the researchers.
The Kiraly model was significantly more accurate than Sybil at predicting lung cancer, with an area under the curve (AUC) of 0.93 versus 0.89. Among Lung-RADS 3+ exams, it also outperformed both the Brock and Sybil models, with an AUC of 0.89 compared with 0.79 and 0.84, respectively.
Attend this Sunday afternoon talk to get all of the details.
