Enhancing Dental Caries Classification with Adversarial Training on Bitewing Radiographs

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate dental caries classification from radiographs is vital for diagnosis and treatment.
• Traditional methods can be subjective; computer-aided diagnostic systems offer potential benefits.
• Deep learning models show promise for caries classification using the ICCMS™ 7-class system.

Purpose of the Study:

• To enhance ResNet models for dental caries classification by incorporating projected gradient descent (PGD) into the training process.
• To improve the robustness and classification performance of deep learning models through adversarial data augmentation.

Main Methods:

• Augmenting a clean dataset with mild perturbations using projected gradient descent (PGD).
• Training ResNet models, specifically ResNet-50, with the PGD-augmented dataset.
• Evaluating model performance using validation and test accuracy, sensitivity, and specificity metrics.

Main Results:

• The PGD-augmented ResNet-50 model demonstrated significant performance improvements.
• Validation accuracy increased from 58.20% to 67.20%.
• Test accuracy improved from 57.14% to 59.18%, with notable gains in sensitivity and specificity.

Conclusions:

• Adversarial training techniques like PGD can substantially improve the accuracy, robustness, and reliability of deep learning models for dental caries classification.
• The findings support the advancement of computer-aided diagnostic tools for clinical dental practice.
• PGD augmentation offers a promising approach to enhance AI-driven dental diagnostics.