Automated assessment of pediatric adenoid hypertrophy on lateral cephalograms using a deep learning–based keypoint approach

Adenoid hypertrophy (AH) is common in children and causes upper airway obstruction, which frequently forces oral breathing. In this study, we developed a deep learning framework (AH_KNet) to automatically identify AH and assess its severity using lateral cephalograms, with the aim of providing an objective, non-invasive aid for screening and referral in routine practice. A cohort of 1,499 lateral cephalograms (899 normal, 600 AH) was collected and divided into training ( n = 1,337) and test ( n = 162) datasets. A validation subset ( n = 162) was randomly extracted from the training set for hyperparameter tuning and internal evaluation. We used a ResNet-50 backbone with a Feature Pyramid Network (FPN) to regress anatomical landmarks and calculate the Adenoidal/Nasopharyngeal (A/N) ratio. To improve robustness, landmark optimization strategy (Method B) was implemented by excluding the unstable Articulare (Ar) point. Statistics included intraclass correlation coefficients (ICCs), mean radial error (MRE), F-value, p -value, accuracy, sensitivity, specificity, Cohen’s kappa, and Receiver Operating Characteristic (ROC) curve analysis. The proposed approach AH_KNet achieved high reliability in landmark detection (peak ICC = 0.991). Compared with the conventional method, performance improved across multiple metrics: accuracy increased from 90.26% to 93.80%, Cohen’s kappa from 69.84% to 80.06%, and Area Under the Curve (AUC) from 0.865 to 0.913. Inference time was under one second per image, supporting efficient integration into clinical workflows. This AI-assisted model offers potential in supporting opportunistic screening pediatric AH. By providing rapid and consistent radiographic assessment, it may assist orthodontists in identifying children who require further otolaryngologic evaluation, while maintaining current diagnostic workflows. The Ethics Committee of the Affiliated Stomatological Hospital of Tongji University ([2024]-SR-56), Date of Registration: Feb.22, 2024.