A deep learning approach to predict the results of root coverage procedures: a retrospective study

Background Gingival recession (GR) occurs frequently in patients with periodontitis. The accuracy of the prediction of the root coverage procedures might be influenced by the periodontists’ subjective bias and the complex anatomical factors of the affected teeth. Deep learning network may improve the accuracy and efficiency of the prediction of the root coverage procedures. This study aims to establish a deep learning model based on intraoral photographs, which could accurately predict the results of root coverage procedures. Methods A total of 963 teeth (408 intraoral photographs) were included to construct the model. A deep learning model based on YOLO-v8 was proposed to annotate the papilla zeniths (PZ), mesial and distal contact points (CP), cemento-enamel junction (CEJ), and point angles (CPA) on each tooth, so that the maximum root coverage (MRC) was predicted by the model. The differences between the manually annotated and the model-predicted key points were calculated to evaluate the accuracy of the model. Another 114 teeth (50 intraoral photographs) 6-month postoperative were included in this study to validate the accuracy of the model on real clinical cases. Results The mean radial error (MRE) for all key points detections was 2.19 pixels, and the standard radial error (SRE) was 3.72 pixels, the MRE between manually annotated and model-predicted key points ranges from 0 to 8 pixels. In real cases, the MRE between the manually annotated GM and the model-predicted MRC was 0.91 pixels, the SRE was 1.32 pixels. Conclusions The model constructed in this study could accurately predict the results of root coverage procedures.