Transfer Learning From Hand-Trained Deep Learning Models to Estimate Bone Age From Knee Radiographs
Orthopaedic Journal of Sports Medicine · 2026-04 · 14 (4)
DOI: 10.1177/23259671261424923 ↗Kurz gefasst
Fachartikel in der Zeitschrift Orthopaedic Journal of Sports Medicine, online erschienen am 24.04.2026. Verfasst von Joshua T Bram und 8 weiteren Autor:innen. Beteiligte Einrichtung: Lerner Children's Pavilion, Hospital for Special Surgery, New York, New York, USA. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: AOSSM Playmaker Grant.
Abstract (Originalsprache)
Background
Orthopaedic surgeons treating skeletally immature patients with knee pathology rely on an accurate evaluation of skeletal age. Although the Greulich and Pyle atlas remains the gold standard for bone age estimation, it requires additional hand imaging and radiation exposure for patients who already have readily available knee imaging.
Purpose
To develop a deep learning (DL) model for bone age estimation based on knee radiographs. Study Design: Cohort study (diagnosis); Level of evidence, 3 Methods: This was an institutional review board-approved study. A total of 2374 cropped knee images—80% for training and 20% for a test set—from patients aged ≤18 years at an orthopaedic specialty hospital with paired hand films obtained within 3 months of knee imaging were used. Patients with previous physeal trauma about the knee, genetic disorders including skeletal dysplasias and endocrinopathies affecting maturation, were excluded. A ConvNeXT model was trained on sex, skeletal age, and chronological age as input values using 5-fold validation. Model performance was evaluated using mean absolute error (MAE) and gradient-based saliency maps to assess model interpretability.
Results
The model achieved an MAE of 5.02 months relative to the assigned ground-truth bone ages, nearly half that of the abbreviated Fels method (9.59 months). An even lower MAE of 3.43 months was observed relative to the pretrained pseudo-labels generated by a hand DL model used to annotate the training set. Bland-Altman analysis revealed near-zero bias, indicating close agreement between predicted and ground truth bone age. This model outperformed both prior simple and artificial intelligence-based skeletal maturity estimation tools of the knee.
Conclusion
This highly accurate DL model for knee skeletal maturity estimation demonstrates that automated bone age estimation from routine knee radiographs is feasible. We anticipate that this model can serve as a valuable diagnostic assistant for orthopaedic surgeons and radiologists evaluating skeletally immature patients and help to augment their clinical decision-making. External validation and model refinement are important for adoption into everyday practice.
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Lerner Children's Pavilion, Hospital for Special Surgery, New York, New York, USA.
- Sports Medicine Institute, Hospital for Special Surgery, New York, New York, USA.
- Department of Computer Engineering, Sharif University of Technology, Tehran, Iran.
- School of Medicine, Tehran University of Medical Sciences, Tehran, Iran.
ORCID-Profile: Joshua T Bram, M Moein Shariatnia, Samuel A Beber, Ruth H Jones, Olivia C Tracey, Daniel W Green, Peter D Fabricant
Förderung
- AOSSM Playmaker Grant
Zitieren
APA
Bram, J. T., Pareek, A., Daliliyazdi, A., Shariatnia, M. M., Beber, S. A., Jones, R. H., Tracey, O. C., Green, D. W., & Fabricant, P. D. (2026). Transfer Learning From Hand-Trained Deep Learning Models to Estimate Bone Age From Knee Radiographs. Orthopaedic Journal of Sports Medicine, 14(4). https://doi.org/10.1177/23259671261424923
Vancouver
Bram JT, Pareek A, Daliliyazdi A, Shariatnia MM, Beber SA, Jones RH, et al. Transfer Learning From Hand-Trained Deep Learning Models to Estimate Bone Age From Knee Radiographs. Orthopaedic Journal of Sports Medicine. 2026;14(4). doi:10.1177/23259671261424923
BibTeX
@article{Bram2026,
author = {Bram, Joshua T and Pareek, Ayoosh and Daliliyazdi, Amir and Shariatnia, M Moein and Beber, Samuel A and Jones, Ruth H and Tracey, Olivia C and Green, Daniel W and Fabricant, Peter D},
title = {Transfer Learning From Hand-Trained Deep Learning Models to Estimate Bone Age From Knee Radiographs},
journal = {Orthopaedic Journal of Sports Medicine},
year = {2026},
volume = {14},
number = {4},
doi = {10.1177/23259671261424923}
}
EUVIMED