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

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Current caries management requires precise diagnosis for minimally invasive strategies.
• Artificial intelligence (AI) tools can enhance caries classification on cone beam computed tomography (CBCT) scans.
• Clinically applicable and interpretable AI solutions for caries detection are needed.

Purpose of the Study:

• To develop and validate CariesAI-3D, an interpretable AI framework for classifying dentin caries severity on CBCT images.
• To ensure accurate and robust performance in caries diagnosis.

Main Methods:

• Developed a multitask learning network, CariesAI-3D, incorporating a spatial-attention feature fusion module (SA-FFM).
• Trained and validated the model on a dataset of 2148 CBCT images, with evaluation using 5-fold cross-validation and an independent test set.
• Assessed performance using accuracy, precision, recall, F1-score, AUC, and mean absolute difference (MAD), comparing against 6 baseline models.

Main Results:

• CariesAI-3D significantly outperformed baseline models, achieving 0.886 accuracy, 0.882 precision, 0.873 recall, and 0.876 F1-score on the cross-validation set.
• The SA-FFM module improved model accuracy, and CariesAI-3D demonstrated strong generalization with class-wise AUCs from 0.947 to 0.998.
• Class activation mapping (CAM) confirmed that model predictions correlated with caries and pulp regions, indicating interpretability.

Conclusions:

• CariesAI-3D accurately and interpretably classifies dentin caries severity on CBCT images.
• The integration of multitask learning and SA-FFM represents a significant advancement over conventional caries diagnosis methods.