Grad-CAM Helps Explainability of Supernumerary Teeth Detection by Deep Learning

Objectives: Supernumerary teeth can occur in any part of the dental arch, but most commonly occurs in the premaxilla. We have been investigating approaches of detecting supernumerary teeth in the early mixed dentition from panoramic radiographic images using Deep Learning. Herein, we examined the applicability of Gradient-weighted Class Activation Mapping (Grad-CAM), a method that visualized the decisions of Deep Learning-based convolutional neural network (CNN) models, to the detection of supernumerary teeth in the early mixed dentition of children.
Methods: We employed three CNN algorithms in this study. The total of 220 panoramic radiograph images were revalidated and diagnosed by two expert pediatric dentists as supernumerary teeth (case group, n = 120) or no anomalies (control group, n = 100). To avoid deviations in the datasets used to train models, a five-fold cross-validation method was employed. The datasets were randomly split into five groups; one group, 20 % of the overall data set, represented the test dataset, while the remaining four groups were used as training dataset. The diagnostic accuracy, precision, recall, F1 score, and area under the curve were calculated for detection and diagnostic performance of the algorithms. Grad-CAM from the final convolution layer was used to visually explain the features. This study was approved by Ethical Committee for Epidemiology of Hiroshima University.
Results: The results of the analysis with the three CNNs showed that the feature visualization using Grad-CAM with heat map identified the location of the supernumerary teeth. The three CNN models achieved high values in each performance metric.
Conclusions: These results suggested that the Grad-CAM approach is potentially applicable to the diagnosis of supernumerary teeth.