IADR Abstract Archives
Automatic Root Resorption Prediction System Based on Deep Learning Method
Objectives: Following orthodontic treatment, external root resorption (ERR) affects more than 90% of patients, threatening teeth function and leading to teeth mobility or even loss. Planning properly and anticipating the effects of ERR will slow down the process. It is necessary to adopt deep learning methods to predict ERR automatically on three-dimensional images. Consequently, the objective of this study are (1) to automatically predict the degree of the posttreatment root resorption after orthodontic treatment.(2) to validate the system’s accuracy on a clinical dental CBCT dataset.
Methods: Four models are adopted in the proposed system. Firstly, For extraction by an effective localization module, an integrated tooth segmentation model based on a combination of multi-scale semantic information with explicit anatomical connection and tooth bounding box is proposed. This model effectively segments teeth into their own category. Secondly, a model that detects tooth landmarks or axes that can be a guide to separate the crown and root. Thirdly, to estimate the crown or root ratios, as well as the change in volume and length of the tooth, a volume and length analysis model will be proposed. Finally, based on patient-specific and treatment-related factors, a prediction model may predict the degree of post-orthodontic root resorption. To validate the system’s performance, we compare our experiment result with the manual measurement result.
Results: The entire protocol took K-Nearest Neighbor for predicting ERR. We used 6 metrics to evaluate the method. The score of recall achieved 0.846 respectively. The accuracy score reached 0.846. The score of precision was 0.716.The F1-measure score was 0.776.
Conclusions: Experiment results demonstrate that the system provides a valuable prediction of ERR from associated factors, which will assist dentists in selecting the best treatment strategy to limit root resorption. In the future, combining different modalities of data may improve the diagnostic accuracy of ERR prediction.
Methods: Four models are adopted in the proposed system. Firstly, For extraction by an effective localization module, an integrated tooth segmentation model based on a combination of multi-scale semantic information with explicit anatomical connection and tooth bounding box is proposed. This model effectively segments teeth into their own category. Secondly, a model that detects tooth landmarks or axes that can be a guide to separate the crown and root. Thirdly, to estimate the crown or root ratios, as well as the change in volume and length of the tooth, a volume and length analysis model will be proposed. Finally, based on patient-specific and treatment-related factors, a prediction model may predict the degree of post-orthodontic root resorption. To validate the system’s performance, we compare our experiment result with the manual measurement result.
Results: The entire protocol took K-Nearest Neighbor for predicting ERR. We used 6 metrics to evaluate the method. The score of recall achieved 0.846 respectively. The accuracy score reached 0.846. The score of precision was 0.716.The F1-measure score was 0.776.
Conclusions: Experiment results demonstrate that the system provides a valuable prediction of ERR from associated factors, which will assist dentists in selecting the best treatment strategy to limit root resorption. In the future, combining different modalities of data may improve the diagnostic accuracy of ERR prediction.
3894185_File000001.jpg