Deep learning algorithm to predict Greulich and Pyle bone age

Toan Nguyen

Deep learning algorithm to predict Greulich and Pyle bone age

Published In

Deep learning algorithm to predict Greulich and Pyle bone age (ESPR2022)

Authors

Aloïs Pourchot ; Vincent Marty ; Jeanne Ventre ; Nor-Eddine Regnard

Abstract

Purpose:
Bone age assessment is a routine procedure in pediatric radiology, for which the Greulich and Pyle atlas [1] is mostly used. The aim of this study is to evaluate the performance of a deep learning algorithm to predict bone age.

Material and Methods:
We collected 162 frontal left and right-hand views of children aged 5 to 19 of which 97 were boys and 65 were girls from an internal existing test set coming from four European radiology departments. Images with a truncated wrist (N=6) preventing bone age estimation were excluded from the dataset. The gold standard was established by a senior board-certified pediatric radiologist who assessed the bone age based on the Greulich and Pyle atlas knowing the sex and chronological age of the patient but blinded to the algorithm results. The deep learning algorithm was engineered to predict the bone age based on a public dataset from RSNA. It was trained on 12,600 frontal left and right-hand views, tested on 200, and validated on 1,000.
The performance of the deep learning algorithm was assessed by the mean average error between the predicted bone age and the Greulich and Pyle bone age. Confidence intervals (CI) were calculated using bootstrap samples. The Pearson coefficient was used to estimate the correlation between predicted and true bone age.

Results:
A total of 156 patients were included in the study of which 93 were boys with mean chronological age 12 ± 3.7 years and 63 were girls with mean chronological age of 9.8 ± 3.6 years.
The mean average error was 0.699 years (95% CI: [0.570, 0.844]) in the boys group with a Pearson correlation coefficient R2 of 0.902. In the girls’ group, the mean average error was 0.572 years (95% CI: [0.458, 0.692]) with a Pearson correlation coefficient of 0.951.

Conclusion:
The deep learning algorithm estimated the Greulich and Pyle bone age with high accuracy.

References
[1] Greulich WW, Pyle SI. Radiographic Atlas of Skeletal Development of the Hand and Wrist, 2nd Edition, Stanford, CA: Stanford University Press and London, UK: Oxford University Press, 1959.