Fabrication of customized dental guide by stereolithography method and evaluation of dimensional accuracy with artificial neural networks

Before dental guide designs are made, careful and detailed treatment planning is very important to ensure a successful and long-lasting treatment. Considering all products and processes that help dental guides in treatment planning are evaluated, the use of technology and software-supported solutions in dental applications is increasing, as in many different areas of life. In the study, the first step of the treatment planning, dental Cone Beam Computed Tomography (CBCT) image data, the elimination of image disorders, the determination of bone and tissue boundaries and the obtaining of a 3D (3D) anatomical model were provided. In line with the results obtained, three (3) different dental guide designs were carried out. Dental guide designs and mandible model were produced with Stereolithography (SLA) 3D printer. Data were obtained from the produced model using a 3D scanner. Dimensional error values were obtained by performing a three-dimensional overlay between the mandible and dental guide models. The obtained dimensional error data were compared statistically between dental guides using the open source Jamovi 2.0.0 statistical software program and the dental guide model with the best placement in the mandible was determined. Artificial Neural Network (ANN) models from artificial intelligence techniques were designed by using the dimensional aperture values according to the coordinates obtained as a result of the registration process. In ANN models, the loss function is used as "binary cross entropy" and the optimization function is used as "man". In addition, "relu" activation functions in the hidden layer of the artificial neural network and "sigmoid" activation functions in the output layer are used. A total of 10 ANN models were designed. In the ANN models, the number of neurons in the 1st layer was 8, the number of neurons in the 2nd layer was 6, the epoch number was 100 and the batch size was 10, and the model gave 99% accuracy.