A clinically oriented and interpretable AI framework for classifying dentin caries severity on CBCT images

Current caries management has emphasized minimally invasive, biologically driven strategies that demand a higher level of precision in caries diagnosis. Artificial intelligence (AI)-driven tools for classifying caries on cone beam computed tomography (CBCT) scans may improve diagnostic accuracy and streamline clinical treatment planning. However, clinically oriented and interpretable AI solutions remain lacking.

The purpose of this study was to develop and validate an interpretable AI framework, CariesAI-3D, for accurate and robust classification of dentin caries severity on CBCT images.

A high-quality CBCT dataset comprising 2148 CBCT images of single teeth was established, including sound teeth, moderate caries, deep caries, and extremely deep caries. The dataset was divided into a 5-fold cross-validation set (1826) for model training and validation and an independent test set (322) for final evaluation. CariesAI-3D was developed as a multitask learning network incorporating a spatial-attention feature fusion module (SA-FFM) for caries classification. Its performance was evaluated against 6 baseline models (ResNet-18, ResNet-34, ResNet-50, DenseNet-121, DenseNet-169, and MobileNet-V2) using cross-validation. An ablation study was conducted to evaluate the effectiveness of the SA-FFM. Caries classification performance was assessed using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic (ROC) curve (AUC). The mean absolute difference (MAD) between cross-validation and independent test sets was calculated to quantify model generalization. Statistical significance was assessed using a corrected resampled t test (α=.05).

CariesAI-3D significantly outperformed the baseline models on the cross-validation set, achieving an accuracy of 0.886, precision of 0.882, recall of 0.873, and F1-score of 0.876. The ablation study confirmed that CariesAI-3D with SA-FFM demonstrated better accuracy than both the backbone model and the model with the element-wise feature addition. Furthermore, CariesAI-3D exhibited strong generalization on the independent test set, achieving class-wise AUC values between 0.947 and 0.998, with metric-wise MAD ranging from 0.011 to 0.033. Class activation mapping (CAM) demonstrated that the model's predictions were highly correlated with caries and pulp regions.

By integrating multitask learning with an SA-FFM, CariesAI-3D achieved the accurate and interpretable classification of dentin caries severity on CBCT images, demonstrating significant advancements over conventional methods.