16 Bit wins the inaugural RSNA machine learning challenge
Chicago. At RSNA 2017, the algorithm submitted by Alexander Bilbily and Mark Cicero placed first in the Radiological Society of North America's first machine learning challenge, which drew 105 submissions from 48 participants, estimating pediatric bone age from hand radiographs with a mean absolute difference of 4.2 months from the reference reading.
Bone age assessment from a hand radiograph is a routine pediatric task with considerable variability between readers. The challenge supplied more than 12,000 labelled hand radiographs for training and scored entries on a separate set of 200.
Each image’s reference bone age combined six expert estimates: the clinical report, four pediatric radiologists, and a repeat read a year later. Individual reviewers differed from the reviewers’ mean by 4.8 to 7.0 months, against the winning algorithm’s 4.2. The challenge report, with Bilbily and Cicero among its authors, was published in Radiology in 2019.
The model became Physis, 16 Bit’s bone age software, licensed by Health Canada in 2022.