Frontal cephalometric landmarking: humans vs artificial neural networks
Frontal cephalometric radiography (frontal ceph) is one of the important diagnostic methods in orthodontics and maxillofacial surgery. It allows one to determine occlusion anomalies in the transverse and vertical planes and to evaluate the symmetry of the facial skeleton relative to the median plane, including analysis of the position of the jawbone. Aim: The aim of this study was to develop an artificial neural network (ANN) for placing cephalometric points (CPs) on frontal cephs and to compare the accuracy of its performance against humans. Materials and methods: The study included 330 depersonalized frontal cephs: 300 cephs for training ANNs and 30 for research. Each image was imported into the ViSurgery software (Skolkovo, Russia) and the 45 CPs were arranged. The CPs were divided into three groups: 1) precise anatomical landmarks; 2) complex anatomical landmarks; and 3) indistinct anatomical landmarks. Two ANNs were used to improve the accuracy of CP placement. The first ANN solved the problem of multiclass image segmentation, and the second regression ANN was used to correct the predictions of the first ANN. The accuracy of CP placement was compared between the ANN and three groups of doctors: expert, regular, and inexperienced. Then, using the Wilcoxon t test, the hypothesis that an ANN makes fewer or as many errors as doctors in the three groups of points was tested. Results: The deviation was estimated by the mean absolute error (MAE). The MAE for the points placed by the ANN, as compared with the control, was close to the average result for the regular doctor group: 2.87 mm (ANN) and 2.85 mm (regular group); 2.47 mm (expert group), and 3.61 mm (inexperienced group). The results for individual groups of points are presented. On average, the ANN places CPs no less accurately than the regular doctor group in each group of points. However, calculating all points in total, this hypothesis was rejected because the P value was 0.0056. A different result was observed among the inexperienced doctor group. Points from groups 2 and 3, as well as all points in total, were placed more accurately by the ANN (P = 0.9998, 0.2628, and 0.9982, respectively). The exception was group 1, where the points were more accurately placed by the inexperienced doctors (P = 0.0006). Conclusion: The results of the present study show that ANNs can achieve accuracy comparable to humans in placing CPs, and in some cases surpass the accuracy of inexperienced doctors (students, residents, graduate students).