Deep Learning–Based Bone Age Assessment for Predicting Final Adult Height in Girls With Central Precocious Puberty
- Author:
Jeong Min SONG
1
;
Pyeong Hwa KIM
;
Young Ah CHO
;
Ah Young JUNG
;
Jin Seong LEE
;
Ja Hye KIM
;
Hee Mang YOON
Author Information
- Publication Type:Original Articl
- From:Korean Journal of Radiology 2026;27(6):568-577
- CountryRepublic of Korea
- Language:English
-
Abstract:
Objective:This study aimed to evaluate the accuracy of predicting final adult height (FAH) in Korean girls with central precocious puberty (CPP) using artificial intelligence (AI)-derived bone age assessments integrated into the Bayley–Pinneau (BP) or Korean National Growth Chart (KGC) prediction models.
Materials and Methods:This single-center, retrospective study included 122 Korean girls with CPP who received gonadotropinreleasing hormone agonist (GnRHa) treatment for at least two years between January 2000 and November 2022. We assessed bone age and predicted adult height at the initiation and completion of GnRHa treatment. We used three bone age assessment methods: human expert assessment based on the Greulich-Pyle (GP) atlas (Human-GP), AI-derived GP (AI-GP), and AI-weighted GP scoring (AI-GPw). We calculated predicted adult heights (PAHs) using both the BP and KGC models, generating 12 PAH estimates per patient (2 time points x 3 bone-age methods x 2 height-prediction models). We assessed prediction accuracy and agreement with FAH using linear regression analysis and Bland–Altman plots and performed multivariable analysis to identify significant predictors of FAH.
Results:Human-GP, AI-GP, and AI-GPw demonstrated comparable overall performance in predicting FAH (R 2 : 0.470–0.646 and 0.691–0.822 for treatment initiation and completion, respectively). AI-GPw combined with BP yielded slightly better point estimates but showed no statistically significant differences. At both time points, the BP model demonstrated consistently narrower 95% limits of agreement (LoA) than the KGC model. Multivariable analysis identified AI-GPw-BP and height percentile score as significant predictors of FAH at both time points; mid-parental height was significant only at treatment initiation.
Conclusion:Human-GP, AI-GP, and AI-GPw demonstrated comparable accuracy in predicting FAH. The BP model demonstrated consistently narrower 95% LoA than did the KGC model. AI-GPw-BP was an independent predictor of FAH. These findings support the clinical utility of AI-derived bone age assessments for individualized FAH prediction in patients with CPP.
