Deep learning–based cephalometric diagnosis of adenoid hypertrophy in children

Accurate anatomical landmark detection is crucial for adenoid hypertrophy assessment using the adenoid-nasopharyngeal (A/N) ratio. Traditional CNN-based regression approaches treat landmark detection as a direct coordinate regression problem, which may not fully capture the spatial relationships between anatomical structures. This study aimed to evaluate YOLO11-Pose, a state-of-the-art pose estimation algorithm, for automatic detection of cephalometric landmarks in adenoid assessment, and to analyze its algorithmic advantages over conventional deep learning approaches. We collected 900 lateral cephalometric radiographs annotated with four landmarks (PNS, A1, Ba, Ar). YOLO11-Pose was compared against six regression-based models (custom CNN, ResNet18, ResNet50, DenseNet121, GoogLeNet, VGG16) and one heatmap baseline (HRNet-W32). The YOLO11 architecture features a CSPDarknet backbone, multi-scale feature fusion through PANet, and a decoupled detection head with integrated pose estimation. YOLO11-Pose achieved the lowest mean pixel error (2.71 ± 1.81 pixels; PCK@0.05 of 99.86%) and was significantly more accurate than the six regression-based models (adjusted p < 0.001 for each comparison; regression pixel error range 3.91–5.72 pixels). At the strictest threshold (PCK@0.01) YOLO11-Pose reached 55.56%, compared with 18.06–31.81% for the other models; the difference was most pronounced at the posterior nasal spine (PNS), where the error was 46.4% lower than that of ResNet18 (2.29 vs 4.27 pixels). In this single-centre validation set, YOLO11-Pose localized the four cephalometric landmarks more accurately than direct coordinate regression and matched a high-resolution heatmap baseline at a small fraction of its cost; its accuracy was comparable to the inter-observer variability of the reference annotations (2.7–3.7 px). Because the individual components were not ablated, the advantage is reported as an empirical property of the detection-and-pose formulation rather than as the proven effect of a single architectural element. External validation and prospective evaluation are still required before the method can be considered for routine clinical adenoid assessment.