Adversarial vulnerability and robustness of deep learning models for panoramic dental X-ray segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Deep learning segmentation is crucial for dental diagnosis and treatment planning.
• Current models are vulnerable to adversarial perturbations, impacting clinical reliability.

Purpose of the Study:

• To systematically study adversarial vulnerability and robustness of deep learning models for panoramic dental X-ray segmentation.
• To develop and evaluate an effective defense strategy against adversarial attacks.

Main Methods:

• Benchmarked 11 model variants on a dataset of 995 panoramic dental X-rays.
• Applied white-box adversarial attacks (FGSM, I-FGSM, PGD, DeepFool) to evaluate model robustness.
• Implemented a customized multi-attack adversarial defense strategy.

Main Results:

• Minimal perturbations caused significant performance drops (IoU collapsed from 0.851 to 0.649 at epsilon=0.0314).
• The defense strategy improved robustness, increasing IoU by 14.9% at epsilon=0.0314 and 12.5% at epsilon=0.0157.
• Defended models produced more stable and anatomically consistent segmentation masks under attack.

Conclusions:

• Deep learning models for dental X-ray segmentation exhibit significant adversarial vulnerability.
• The proposed defense strategy enhances model robustness, crucial for safety-critical clinical applications.
• This study establishes a benchmark for adversarial robustness in dental image segmentation.