Determination of cephalometric points with artificial intelligence algorithms developed by deep learning on lateral cephalometric images obtained from conic-beam computer tomography images

Aim: Today, new technologies have been developed based on the principles of imitating the functioning of the human brain. These are mainly based on artificial intelligence (AI) technologies. The aim of this research is to investigate the success of cephalometric landmark detection using a special artificial intelligence (AI) algorithm. In our study, it is aimed to minimize the errors made by using a special artificial intelligence algorithm for the differences between observers that may occur in the determination of orthodontic landmarks and the inconsistencies that can be observed in the evaluations of the observers at different times. Material and Method: In our study, whose data set consists of cephalometric radiograph images obtained from CBCT images of patients before treatment, the success of Pytorch architecture in determining the skeletal, dental and soft tissue points used in the determination of orthodontic measurements on radiographs was compared. Labeling of cephalometric points on images was done using CranioCatch labeling software (CranioCatch, Eskişehir, Turkey). The training of the AI model was carried out using 300 Epochs with the CNN-based deep learning method applied to PyTorch. The learning rate of the model was determined as 0.0001. Results: The highest SDR score of 2 mm, 2.5 mm, 3 mm and 4 mm was obtained from Columella (Cm), Mandibular 1st molar mesial (Md6m) and Soft tissue B(B') points as 100, 100, 100 and 100, respectively. The MRE ± SD of Cm was 0.78 ± 0.41, the MRE±SD of Md6m was 0.74±0.38, and the MRE ± SD of B was 0.72±0.38. The lowest SDR score was obtained from the 2 mm, 2.5 mm, 3 mm and 4 mm Pterygoid(Pt) points as 65.2, 77.8, 85.2 and 94.7, respectively. MRE ± SD of Pt was found to be 1.731±1.17. Conclusion: Our study is very important in terms of the development of CBCT reporting programs with deep learning infrastructure to be made in the future. We anticipate that these programs will have a role in decision support system by saving time for physicians in routine clinical practices.