Validation of BoneXpert bone age across five US centers
A Stanford-led study, with contributing centers in Boston, Atlanta, Cincinnati, Yale and Stanford, has been published in Pediatric Radiology.
Hand radiographs of 1,285 children were each rated four times by a panel of 22 radiologists, and also assessed by two automated methods: BoneXpert and a deep learning AI developed at Stanford.
The design enabled an interchangeability study, where a reference rating was formed by averaging three of the manual ratings, so that BoneXpert, the deep learning method, and the remaining human rating could each be compared against the same reference.
Measured this way, BoneXpert came closest to the reference — with the lowest mean absolute error and the fewest large deviations of the three:
|
|
BoneXpert |
Deep Learning |
Human raters |
|
Mean absolute error |
4.8 months |
6.3 months |
6.5 months |
|
Rate of deviations > 1.8 years |
0.7 % |
2.4 % |
3.4% |
