Identifying Buccal Furcation Involvement via Artificial Intelligence on Periapical Radiographs

Objectives: Periodontitis is an inflammatory disease causes bone loss around teeth. Early diagnosis and treatment are important to prevent tooth loss from periodontitis. However, it is challenging to detect clinical periodontal furcation involvement. Convolutional neural networks (CNN), currently used for deep learning in computer vision¹, have the potential to be applied to assist dental radiographic diagnosis. Therefore, the objective of this study is to determine effectiveness of applying CNNs to identify buccal/lingual furcation involvements in periapical radiographs.
Methods: Full mouth radiographs of 139 patients were included. 121 of the total 19,064 radiographs were identified with buccal/lingual furcation involvement by a periodontist. Two examiners were calibrated through 3 sessions for 5 hours (calibration included identification of different radiolucent features) before manually labelling each furcation involvement for CNN training. Furcation involvements were labelled using MS paint online² (Figure 1). The labeled radiographs were subsequently divided into 2 datasets: 101 for training, 20 for testing. These images were preprocessed, and the CNN model was created with the application of U-Net³. Test sensitivity, specificity, and positive and negative predictive value of the CNN were calculated (Table 1).
Results: Within the 20 radiographs in the testing set, 37 molars were identified and the results of CNN recognition of furcation defects were 54.1% (n=20) true positive (TP), 8.1% (n=3) false positive (FP), 10.8% (n=4) false negative (FN), and 27% (n=10) true negative (TN). The sensitivity of the CNN identification of furcation was 0.83 and the specificity was 0.77. The positive predictive value was 0.87 and the negative predictive value was 0.71 (Table 2).
Conclusions: The CNN was able to identify buccal/lingual furcation involvements. Interestingly, we’ve noticed instances where CNN were able to correct human errors (Figure 2). However, it is clear that further developing machine learning systems is necessary for more reliable assessments.