Deep Learning for the Radiographic Detection of Periodontal Bone Loss.

We applied deep convolutional neural networks (CNNs) to detect periodontal bone loss (PBL) on panoramic dental radiographs. We synthesized a set of 2001 image segments from panoramic radiographs. Our reference test was the measured % of PBL. A deep feed-forward CNN was trained and validated via 10-times repeated group shuffling. Model architectures and hyperparameters were tuned using grid search. The final model was a seven-layer deep neural network, parameterized by a total number of 4,299,651 weights. For comparison, six dentists assessed the image segments for PBL. Averaged over 10 validation folds the mean (SD) classification accuracy of the CNN was 0.81 (0.02). Mean (SD) sensitivity and specificity were 0.81 (0.04), 0.81 (0.05), respectively. The mean (SD) accuracy of the dentists was 0.76 (0.06), but the CNN was not statistically significant superior compared to the examiners (p = 0.067/t-test). Mean sensitivity and specificity of the dentists was 0.92 (0.02) and 0.63 (0.14), respectively. A CNN trained on a limited amount of radiographic image segments showed at least similar discrimination ability as dentists for assessing PBL on panoramic radiographs. Dentists' diagnostic efforts when using radiographs may be reduced by applying machine-learning based technologies.

Deep Learning for the Radiographic Detection of Apical Lesions.

INTRODUCTION: We applied deep convolutional neural networks (CNNs) to detect apical lesions (ALs) on panoramic dental radiographs. METHODS: Based on a synthesized data set of 2001 tooth segments from panoramic radiographs, a custom-made 7-layer deep neural network, parameterized by a total number of 4,299,651 weights, was trained and validated via 10 times repeated group shuffling. Hyperparameters were tuned using a grid search. Our reference test was the majority vote of 6 independent examiners who detected ALs on an ordinal scale (0, no AL; 1, widened periodontal ligament, uncertain AL; 2, clearly detectable lesion, certain AL). Metrics were the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and positive/negative predictive values. Subgroup analysis for tooth types was performed, and different margins of agreement of the reference test were applied (base case: 2; sensitivity analysis: 6). RESULTS: The mean (standard deviation) tooth level prevalence of both uncertain and certain ALs was 0.16 (0.03) in the base case. The AUC of the CNN was 0.85 (0.04). Sensitivity and specificity were 0.65 (0.12) and 0.87 (0.04,) respectively. The resulting positive predictive value was 0.49 (0.10), and the negative predictive value was 0.93 (0.03). In molars, sensitivity was significantly higher than in other tooth types, whereas specificity was lower. When only certain ALs were assessed, the AUC was 0.89 (0.04). Increasing the margin of agreement to 6 significantly increased the AUC to 0.95 (0.02), mainly because the sensitivity increased to 0.74 (0.19). CONCLUSIONS: A moderately deep CNN trained on a limited amount of image data showed satisfying discriminatory ability to detect ALs on panoramic radiographs.