Artificial intelligence detection of external root resorption in teeth using 3D radioconjugated jaw model

Aim: It is the development of automatic external root resorption detection algorithms by placing the extracted human teeth with external root resorption, created by chemical means, into the mandibular bone model obtained with a 3 dimesional printer radiographically compatible with the density of the jawbones, and then introducing the radiographs obtained from this model to the artificial intelligence program. Material and Method: 110 extracted teeth; It was demineralized by applying a 40% nitric acid solution for 8 hours, 8% sodium hypochlorite for 10 minutes, and then washing with distilled water. In the preparation of the radioconjugate phantom model where the teeth will be placed, gray scale values were calculated by comparing the use of calcium sulfate dihydrate, barium sulfate and hydroxyapatite materials. The data set obtained from the teeth used in the study consists of 584 periapical radiographs in total. Resorption areas were created by the polygonal drawing method using CranioCatch labeling software (CranioCatch, Eskişehir, Turkey). Results: When the mass-gray scale values of calcium sulfate dihydrate, barium sulfate, and hydroxyapatite are compared; The average gray value of barium sulfate was found to be 255 and this value shows that it has a very high radiopacity compared to the other 2 materials. The gray scale values of calcium sulfate dihydrate range from about 122 to 241. Although these values are more acceptable when compared to the radiopacity of barium sulfate, the R2 value of 0.96 means that it exhibits less linearity compared to hydroxyapatite. The use of hydroxyapatite as a radiopacity enhancing material was preferred in our study, since hydroxyapatite shows an R2 value very close to the calibration blocks and the minimum gray value is much lower than the other 2 materials. When the sensitivity, precision and F1 score values of the YOLOv5x-cls model for EKR detection are examined, the scores being 1.0 indicates that the model has a high success rate in the testing phase. When the F1 score values of the YOLOv5x-seg model are examined, the fact that the F1 score is 0.8593 indicates that the model works effectively during the testing phase, but it is seen that the classification is much more successful than the segmentation model. Conclusion: As a result of this study, a phantom model compatible with jawbone radiopacity was developed and used in the radiological evaluation of teeth with external root resorption. High success rates have been achieved in detecting external root resorptions observed on radiographs using artificial intelligence.