Comparative Study Of Automatic Landmarking & Diagnostic Models On Cephalograms

With the application of deep learning technique, the automatic cephalometric analysis has been provided. Although automatic landmarking models for cephalometric analysis have been developed, their accuracy still requires validation, and they depend on the expertise of clinicians to resolve discrepancies between the values. To overcome the problem, the automatic diagnosis model has been noticed. However, there is no direct evidence of the superiority of these two modes. In addition, due to the samples studied in most research from a single center, the generalization ability and transfer ability between the two models need to be studied. Based on the same northern Chinese population test set data and the data of the IEEE (Institute of Electrical and Electronics Engineers) 2015 ISBI (International Symposium on Biomedical Imaging) Grand Challenge dataset, we compared the performance, generalization ability, and transfer ability of the proposed two models, respectively. It showed that the automatic landmarking model outperformed the automatic diagnostic model in both test datasets, with a particularly notable accuracy of 90.80% on the IEEE dataset. These findings suggest that the automatic diagnosis model can effectively overcome the limitations of the landmarking mode, but the landmarking measurement is still needed as a reference, and the generalization and transfer ability to other populations need to be further improved.