Detection of cephalometric landmarks with artificial intelligence methods

Cephalometric analyses, conducted using points determined on cephalometric images, are a crucial tool for identifying orthodontic problems and determining treatment methods. Traditional methods of analysis are time-consuming and the results can vary depending on the expertise of the specialist performing the analysis. The aim of this study is to investigate the success of artificial intelligence methods in detecting cephalometric analysis points and whether they can assist specialist practitioners. In the study, the publicly available dataset released for the development of cephalometric point detection algorithms as part of The IEEE ISBI 2015 Challenge in 2015 was used. This dataset consists of a total of 400 lateral cephalometric images, each marked with 19 different points. Of these images, 40 were used for testing and 360 for training. A CNN network was developed using the Pytorch library with the Python programming language for model training. Training was conducted for 300 epochs with a learning rate of 0.001, and the model was saved at the end of the training. In clinical practice, an acceptable margin of error for cephalometric point detection is 2 millimeters. Predictions exceeding this value are considered to potentially lead to incorrect problem identification and treatment planning. As a result of the training, it was observed that the most successfully detected point was the Incision superius point with values of 7.82049 MRE and 0.975 SDR, while the least successfully detected point was the Porion point with values of 43.20057 MRE and 0.225 SDR. In terms of overall success, a successful detection rate of %76.84 was achieved. The training conducted with the existing dataset shows that the detection of cephalometric points using artificial intelligence methods approaches a level of success that can serve as a decision support mechanism for physicians. If the study is further developed and made available to physicians, it is believed that it can provide significant time savings and minimize inconsistencies between labels made by different physicians or by the same physicians at different times