In-House AI-Assisted Cephalometric Landmark Tracing Improves Senior Dental Students’ Accuracy: A Counterbalanced Crossover Study

Background: Cephalometric landmark localisation remains a challenging skill for dental students because it requires radiographic interpretation, anatomical judgement, and accurate translation of landmark definitions into image coordinates. Artificial intelligence (AI)-assisted landmarking may support this process by providing a scaffolded starting point while preserving learner verification and correction. This study evaluated whether an in-house AI-assisted cephalometric tracing system improved landmark localisation accuracy among senior undergraduate dental students.

Methods: A counterbalanced crossover comparative validation study was conducted with 24 senior undergraduate dental students at a dental teaching hospital in Muscat, Oman, from February to May 2026. Students traced 11 de-identified lateral cephalograms under two conditions: unaided manual tracing and AI-assisted tracing. The AI system was developed locally using a deep-learning heatmap-regression workflow and trained and internally evaluated on 499 cephalograms, comprising 399 training, 50 validation, and 50 testing images. Eight anatomical landmarks were analysed: S, N, A, B, UIT, UIA, LIT, and LIA. Student coordinates were compared with a three-specialist reference standard generated by averaging the specialists’ coordinates. Mean radial error, median error, and success detection rates within 2 mm and 4 mm were calculated.

Results: AI-assisted tracing substantially reduced localisation error compared with unaided manual tracing. Across 2,112 landmark observations per condition, mean error decreased from 3.169 mm to 1.826 mm, while median error decreased from 2.320 mm to 1.733 mm. The success detection rate within 2 mm increased by 20.4 percentage points, from 42.5% to 62.9%, and the success detection rate within 4 mm increased by 33.1 percentage points, from 64.0% to 97.1%. The paired student-level comparison favoured AI assistance (Wilcoxon W = 24, P < 0.001), with a mean paired reduction of 1.343 mm (95% CI: 0.739-1.947) and Cohen’s dz = 0.939. AI-assisted mean error was lower for the four skeletal landmarks but not for the four dental landmarks. No statistically significant difference in mean reduction was found between tracing sequences (P = 0.894).

Conclusion: AI-assisted cephalometric tracing improved overall landmark localisation accuracy among senior undergraduate dental students, although the benefit varied by landmark. The findings support a human-AI collaboration model in which AI provides rapid landmark suggestions while students retain responsibility for inspection, adjustment, and final landmark placement. Within the conditions tested, the locally developed system showed potential as an educational scaffold for cephalometric training.