Classification of Caries Based on CBCT: A Deep Learning Network Interpretability Study.

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Dental caries diagnosis relies on visual inspection and radiography, which have limitations.
• Cone-beam computed tomography (CBCT) offers detailed 3D imaging for dental diagnostics.
• Deep learning models show promise in enhancing diagnostic accuracy in various medical fields.

Purpose of the Study:

• To develop a caries classification scheme using CBCT.
• To create and evaluate two deep learning models for improved caries classification accuracy.
• To introduce a metric for treatment strategy determination in type II caries.

Main Methods:

• A dataset of 2713 axial slices from 204 carious teeth CBCT images was curated.
• Two deep learning models (direct and interpretable) were trained using pretrained networks (ResNet50_vd, MobileNetV3_large_ssld).
• The Local Interpretable Model-agnostic Explanations (LIME) method was used for model interpretability analysis.

Main Results:

• The direct classification model achieved a maximum accuracy of 0.700.
• The interpretable classification model consistently exceeded 0.917 in performance metrics.
• LIME confirmed model interpretability by highlighting key image features.
• A significant negative correlation was found between caries-pulp distance and treatment strategy.

Conclusions:

• The CBCT-based caries classification scheme and deep learning models are effective tools for dental caries diagnosis.
• The interpretable deep learning model significantly outperforms the direct model in caries classification.
• The proposed caries classification and models can aid in clinical decision-making for dental caries treatment.