Answers are grounded in this paper and cite their sources.
AI-based cephalometric landmark detection is usually evaluated with localization metrics only. This study traced uncertainty in a YOLO-based pipeline from landmark annotation to ANB-based skeletal classification. Fourteen YOLO configurations with bounding boxes of 40–150 px were trained on augmented samples from 120 lateral cephalograms. The models were evaluated on 11 test radiographs annotated by 22 orthodontists in two sessions. Systematic error was assessed by the directional coherence of landmark offsets. Propagation to the ANB angle was assessed by Monte Carlo simulation with empirical anisotropic error covariances and Sobol sensitivity analysis. All models showed a coherent offset (R = 0.983). Its magnitude increased with box size with a slope of 0.708. This value matches the 0.707 expected for coordinates exported from the box corner. A fixed shift in half the box size corrected the offset without use of test data. After correction, the best model reached a mean radial error of 1.26 mm (95% CI 0.96–1.61 mm). Larger boxes performed worse only on low-resolution radiographs. A-point (S1 = 0.741) dominated ANB variance, followed by B-point (0.157) and N-point (0.085). The ANB limits of agreement of the 40 px models were −2.05° to +1.44°. Skeletal-class concordance with the consensus was 96.9% for the models and 96.5% for individual observers. A single observer deviated from an independent reference by 1.33 mm. The best model deviated by 1.30 mm. Pipeline bias checks should precede benchmarking. A-point localization is the priority target for improving ANB-based diagnosis.
No takes yet. Share an insight, caveat, or question.
Kotuła et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: