Abstract:

Purpose: This study aims to examine the diagnostic performance of detecting pulp stones with a deep learning model on bite-wing radiographs. Material and Methods: 2203 radiographs were scanned retrospectively. 1745 pulp stones were marked on 1269 bite-wing radiographs with the CranioCatch labeling program (CranioCatch, Eskişehir, Turkey) in patients over 16 years old after the consensus of two experts of Maxillofacial Radiologists. This dataset was divided into 3 grou as training (n = 1017 (1396 labels), validation (n = 126 (174 labels)) and test (n = 126) (175 labels) sets, respectively. The deep learning model was developed using Mask R-CNN architecture. A confusion matrix was used to evaluate the success of the model. Results: The results of precision, sensitivity, and F1 obtained using the Mask R-CNN architecture in the test dataset were found to be 0.9115, 0.8879, and 0.8995, respectively. Discussion- Conclusion: Deep learning algorithms can detect pulp stones. With this, clinicians can use software systems based on artificial intelligence as a diagnostic support system. Mask R-CNN architecture can be used for pulp stone detection with approximately 90% sensitivity. The larger data sets increase the accuracy of deep learning systems. More studies are needed to increase the success rates of deep learning models.

Keywords: