Beyond chronological age: uncovering disease-dependent variability in AI-driven pediatric bone age assessment

Abstract Objectives To assess performance and agreement of a commercial artificial intelligence (AI) tool relative to routine pediatric radiologists’ bone age assessments in a large pediatric cohort, with focus on different groups of growth disorders. Materials and methods One thousand five hundred sixteen anterior-posterior hand radiographs of pediatric patients (50.6% boys, mean age 10.84 ± 3.66 years) were retrospectively analyzed using the AI tool BoneAge. AI results were compared with board-certified pediatric radiologists’ reports using Bland–Altman plots, Spearman correlations, and the intraclass correlation coefficient (ICC). Subgroup analyses were performed for three major indications: genetic/neuroendocrine ( n = 1179), orthopedic ( n = 152), and oncologic disorders ( n = 185). Results Mean absolute difference (MAD) in estimated bone age by radiologists and AI was 6.4 ± 7.0 months for the whole cohort, with complete concordance in 45.1% of cases. Best results were achieved in orthopedic questions with an MAD of 4.1 ± 5.6 months (genetic/neuroendocrine: 6.6 ± 7.2, oncologic: 6.7 ± 6.3 months). There was significant correlation between AI and radiologists for the whole study population (ρ = 0.980), and the three main groups: genetic/neuroendocrine (ρ = 0.974), orthopedic (ρ = 0.952), and oncologic indications (ρ = 0.988). ICCs for the entire cohort and the main indication groups were > 0.9, indicating excellent agreement between radiologist and AI. Conclusion The investigated algorithm showed good agreement with bone age assessments performed by pediatric radiologists in a large pediatric cohort. Novel to previous studies, the indication for imaging was considered, demonstrating that agreement between algorithm output and radiologists’ assessments varied across disease groups, with highest agreement observed in orthopedic indications. Key Points Question The performance of artificial intelligence tools for bone age assessment among different subgroups of pediatric patients according to underlying diseases has not yet been investigated. Findings Automated bone age assessment reliably correlates with estimation by pediatric radiologists. Software output differs according to disease groups, with best results for orthopedic indications. Critical relevance statement In a real-world scenario, the commercial software showed good agreement with routine bone age assessments in a large cohort of pediatric patients. The observed disease-dependent variation suggests that radiologists should consider disease-specific differences when using additional automated tools.

Authors

Institutions

Publication Details

Journal
Insights into Imaging
Published
2026-11-14
DOI
https://doi.org/10.1186/s13244-026-02477-0
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Beyond chronological age: uncovering disease-dependent variability in AI-driven pediatric bone age assessment

Sophia F. U. Blum, Ralf‐Thorsten Hoffmann, Marie‐Luise Kromrey, Felix Schön et al.
Insights into Imaging
Artificial Intelligence in Healthcare and Education
article

Beyond chronological age: uncovering disease-dependent variability in AI-driven pediatric bone age assessment

Sophia F. U. Blum, Ralf‐Thorsten Hoffmann, Marie‐Luise Kromrey, Felix Schön, Stephanie Spieth, Jens-Peter Kühn, Christoph Meißner, Kathleen Pätzold
article en

Abstract

Abstract Objectives To assess performance and agreement of a commercial artificial intelligence (AI) tool relative to routine pediatric radiologists’ bone age assessments in a large pediatric cohort, with focus on different groups of growth disorders. Materials and methods One thousand five hundred sixteen anterior-posterior hand radiographs of pediatric patients (50.6% boys, mean age 10.84 ± 3.66 years) were retrospectively analyzed using the AI tool BoneAge. AI results were compared with board-certified pediatric radiologists’ reports using Bland–Altman plots, Spearman correlations, and the intraclass correlation coefficient (ICC). Subgroup analyses were performed for three major indications: genetic/neuroendocrine ( n = 1179), orthopedic ( n = 152), and oncologic disorders ( n = 185). Results Mean absolute difference (MAD) in estimated bone age by radiologists and AI was 6.4 ± 7.0 months for the whole cohort, with complete concordance in 45.1% of cases. Best results were achieved in orthopedic questions with an MAD of 4.1 ± 5.6 months (genetic/neuroendocrine: 6.6 ± 7.2, oncologic: 6.7 ± 6.3 months). There was significant correlation between AI and radiologists for the whole study population (ρ = 0.980), and the three main groups: genetic/neuroendocrine (ρ = 0.974), orthopedic (ρ = 0.952), and oncologic indications (ρ = 0.988). ICCs for the entire cohort and the main indication groups were > 0.9, indicating excellent agreement between radiologist and AI. Conclusion The investigated algorithm showed good agreement with bone age assessments performed by pediatric radiologists in a large pediatric cohort. Novel to previous studies, the indication for imaging was considered, demonstrating that agreement between algorithm output and radiologists’ assessments varied across disease groups, with highest agreement observed in orthopedic indications. Key Points Question The performance of artificial intelligence tools for bone age assessment among different subgroups of pediatric patients according to underlying diseases has not yet been investigated. Findings Automated bone age assessment reliably correlates with estimation by pediatric radiologists. Software output differs according to disease groups, with best results for orthopedic indications. Critical relevance statement In a real-world scenario, the commercial software showed good agreement with routine bone age assessments in a large cohort of pediatric patients. The observed disease-dependent variation suggests that radiologists should consider disease-specific differences when using additional automated tools.

Insights into ImagingVol. 17(1)
University Hospital Carl Gustav Carus (DE), Technische Universität Dresden (DE)
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon