CNN-based remote dental diagnosis model for caries detection with grad-CAM

Abstract

Dental caries is a prevalent global condition, and its diagnosis often requires direct clinical examination by a dentist. However, access to traditional dental care can be limited due to high costs, availability, and patient discomfort. To address these limitations, this study introduced a remote caries detection model using a ResBlock-AutoEncoder that generates domain-specific pre-trained weights. The model demonstrated exceptional performance, achieving an accuracy of 0.9989, an F1-score of 0.9979, and a precision of 1.0, while maintaining a low average inference time of 5.7939 seconds. Furthermore, Grad-CAM was employed to enhance interpretability by visually localizing caries, ensuring model reliability. Notably, this high precision is attributed to the specific characteristics of frontal oral images, which allow for clearer visibility of caries compared to other imaging angles. However, this also introduces a potential limitation, as it does not account for variability in other perspectives of oral images. To improve generalization, future research will incorporate multi-angle dental images.

Keywords: Convolutional neural network; Dental caries; Explainable artificial intelligence; Gradient-weighted class activation mapping; Oral imaging.

Fig. 1

Confusion matrix.

Fig. 2

Train and validation loss of proposed RB-AE model.

Fig. 3

Training and validation results: loss and accuracy for both datasets.

Fig. 4

confusion matrix.

Fig. 5

False positive misclassification cases.

Fig. 6

Grad-CAM visualization for TP cases in the lower and upper jaw.

Fig. 7

Comparison between Grad-CAM and ground truth label.

Fig. 8

Architecture of the proposed ResFC model based on RB-AE.

Fig. 9

Overall structure of class activation mapping.

Fig. 10

Data distribution and data split.

Fig. 11

Image preprocessing.