% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@ARTICLE{Rassmann:281434,
author = {Rassmann, Sebastian and Abashishvili, Luka and Melikidze,
Elene and Sukhiashvili, Anastasia and Lartsuliani, Megi and
Chkhaidze, Ivane and Tskhvediani, Nino and Gordeziani,
Tinatin and Kvaratskhelia, Ekaterine and Kheladze, Nino and
Rekhviashvili, Maia and Rodonaia, Salome and Sukhitashvili,
Natia and Urushadze, Nata and Krawitz, Peter and Tkemaladze,
Tinatin and Javanmardi, Behnam},
title = {{P}opulation-specific calibration and validation of an
open-source bone age {AI}.},
journal = {Scientific reports},
volume = {15},
number = {1},
issn = {2045-2322},
address = {[London]},
publisher = {Springer Nature},
reportid = {DZNE-2025-01119},
pages = {32673},
year = {2025},
abstract = {Assessing skeletal maturity through bone age (BA)
evaluation is crucial for monitoring children's growth and
guiding treatments, such as hormonal therapy and orthopedic
interventions. In recent years, artificial intelligence (AI)
methods have been developed to automate BA assessment.
However, bone growth patterns may vary by ancestry, and many
AI models are trained on limited population datasets,
raising concerns about their applicability to populations
not included in the training process. To address this
shortcoming for the case of the Georgian population, we
retrospectively collected 381 pediatric hand X-rays and
established a manual BA reference rating from seven local
pediatric radiologists and endocrinologists. We then used a
subset of 121 images to perform a sex-specific linear
calibration of the open-source AI, Deeplasia, creating
Deeplasia-GE. On the held-out test set (n = 260), the
default version of Deeplasia achieved a mean absolute
difference (MAD) of 6.57 months, which improved to 5.69
months after calibration. We observed that the default
Deeplasia overestimates the BA in the Georgian cohort with a
signed mean difference (SMD) of + 2.85 and + 5.35 months for
females and males respectively, which after calibration is
significantly reduced to -0.03 and + 0.58 months for females
and males, respectively. We find that Deeplasia-GE has a
smaller error than all the raters and, by design,
Deeplasia-GE inherits the high test-retest reliability from
Deeplasia. These findings suggest that Deeplasia-GE is a
reliable AI-based BA assessment method for Georgian
children.},
keywords = {Humans / Male / Female / Age Determination by Skeleton:
methods / Child / Artificial Intelligence / Calibration /
Retrospective Studies / Child, Preschool / Adolescent /
Infant / Reproducibility of Results / Artificial
intelligence (Other) / Global health equity (Other) / Hand
x-rays (Other) / Model calibration (Other) / Open-Source
(Other) / Pediatric bone age (Other)},
cin = {AG Reuter},
ddc = {600},
cid = {I:(DE-2719)1040310},
pnm = {354 - Disease Prevention and Healthy Aging (POF4-354)},
pid = {G:(DE-HGF)POF4-354},
typ = {PUB:(DE-HGF)16},
pubmed = {pmid:40987805},
doi = {10.1038/s41598-025-20148-w},
url = {https://pub.dzne.de/record/281434},
}