Deep learning for the early identification of periodontitis: a retrospective, multicentre study

•Developing convolutional neural network model for periodontitis using radiographs. •Regions of interest predicted by the model are periodontitis bone lesions. •Ability of the model for periodontitis reaches the level of periodontal experts. •The time required to read each radiograph by the model was shorter than clinicians. Aim To develop a deep-learning model to help general dental practitioners diagnose periodontitis accurately and at an early stage. Materials and methods First, the panoramic radiographs (PARs) from the Second Affiliated Hospital of Nanchang University were input into the convolutional neural network (CNN) architecture to establish the PAR-CNN model for healthy controls and periodontitis patients. Then, the PARs from the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine were included in the second testing set to validate the effectiveness of the model with data from two centres. Heat maps were produced using a gradient-weighted class activation mapping method to visualise the regions of interest of the model. The accuracy and time required to read the PARs were compared between the model, periodontal experts, and general dental practitioners. Areas under the receiver operating characteristic curve (AUCs) were used to evaluate the performance of the model. Results The AUC of the PAR-CNN model was 0.843, and the AUC of the second test set was 0.793. The heat map showed that the regions of interest predicted by the model were periodontitis bone lesions. The accuracy of the model, periodontal experts, and general dental practitioners was 0.800, 0.813, and 0.693, respectively. The time required to read each PAR by periodontal experts (6.042 ± 1.148 seconds) and general dental practitioners (13.105 ± 3.153 seconds), which was significantly longer than the time required by the model (0.027 ± 0.002 seconds). Conclusion The ability of the CNN model to diagnose periodontitis approached the level of periodontal experts. Deep-learning methods can assist general dental practitioners to diagnose periodontitis quickly and accurately. To develop a deep-learning model to help general dental practitioners diagnose periodontitis accurately and at an early stage. First, the panoramic radiographs (PARs) from the Second Affiliated Hospital of Nanchang University were input into the convolutional neural network (CNN) architecture to establish the PAR-CNN model for healthy controls and periodontitis patients. Then, the PARs from the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine were included in the second testing set to validate the effectiveness of the model with data from two centres. Heat maps were produced using a gradient-weighted class activation mapping method to visualise the regions of interest of the model. The accuracy and time required to read the PARs were compared between the model, periodontal experts, and general dental practitioners. Areas under the receiver operating characteristic curve (AUCs) were used to evaluate the performance of the model. The AUC of the PAR-CNN model was 0.843, and the AUC of the second test set was 0.793. The heat map showed that the regions of interest predicted by the model were periodontitis bone lesions. The accuracy of the model, periodontal experts, and general dental practitioners was 0.800, 0.813, and 0.693, respectively. The time required to read each PAR by periodontal experts (6.042 ± 1.148 seconds) and general dental practitioners (13.105 ± 3.153 seconds), which was significantly longer than the time required by the model (0.027 ± 0.002 seconds). The ability of the CNN model to diagnose periodontitis approached the level of periodontal experts. Deep-learning methods can assist general dental practitioners to diagnose periodontitis quickly and accurately.