Automated cephalometric landmark detection is usually evaluated on a single dataset. This study evaluated a dual-encoder regression network across devices, datasets and patient ages.

In this retrospective diagnostic accuracy study, separate models were trained and tested within Aariz (1000 cephalograms, seven devices, 26 landmarks), ISBI 2015 (400 cephalograms, 19 landmarks) and CephAdoAdu (10 landmarks; public validation subset of 300 adolescent and adult images), and each model was applied to the other datasets. Outcomes were the mean radial error (MRE; nominal millimetres for CephAdoAdu) and the success detection rate within 2 mm (SDR). Adolescents and adults were compared by equivalence testing (margin ± 0.25 mm).

On ISBI 2015 Test2, which was not used for model selection, the MRE was 1.46 mm (95% CI 1.42–1.51; SDR 74.7%); it was 1.10 mm (1.01–1.19; 86.9%) on Aariz and 1.01 mm (0.97–1.05; 89.7%) on CephAdoAdu. On Aariz, the MRE ranged from 0.42 to 1.99 mm across devices. Adolescents and adults differed by −0.08 mm (90% CI −0.14 to −0.01; equivalent). Applied to another dataset, the MRE on shared landmarks increased in five of six combinations (by 34–360%). Removing the refinement encoder raised the MRE to 12.6 mm, and a ResNet-34 backbone was at least as accurate as ConvNeXt V2.

Mean errors were close to 1 mm within each dataset and equivalent in adolescents and adults, but accuracy depended on the device, was lowest for posterior landmarks and usually fell on another dataset. Posterior landmarks should be reviewed and accuracy verified on local radiographs before clinical use.