Unsupervised Metal Artifact Reduction in Dental CBCT using Fine-tuned Cycle-Consistent Adversarial Networks

This study addresses the persistent challenge of metal artifacts in dental cone-beam computed tomography (CBCT), which often obscure critical anatomical structures like alveolar bone and root canals. Traditional metal artifact reduction (MAR) methods typically rely on either complex physical modeling or supervised deep learning, the latter of which requires voxel-aligned paired datasets that are ethically and technically difficult to obtain in clinical practice. To overcome this, the authors propose an unsupervised framework using a Cycle-Consistent Adversarial Network (CycleGAN).

The researchers utilized approximately 4,000 images from the public ToothFairy dataset, split into unpaired artifact-affected and artifact-free cohorts. The architecture employs U-Net-based generators to capture multi-scale anatomical features and PatchGAN discriminators to enforce local texture realism. By carefully tuning loss functions—specifically incorporating cycle-consistency and identity losses—the model learns to suppress streak and shading artifacts while minimizing the risk of generative hallucinations, which are common in unsupervised GAN-based image translation.

Quantitative evaluation on a held-out test set showed significant improvements in image quality, including a 34.6% increase in the BRISQUE score and a superior Structural Similarity Index Measure (SSIM) of 0.9105. The model also demonstrated a substantial reduction in the Fréchet Inception Distance (FID) from 207.03 to 157.04. Crucially, the framework achieves a real-time inference speed of 3.03 ms per slice, making it highly suitable for integration into existing digital dentistry workflows without requiring hardware modifications or manual preprocessing.

By eliminating the need for paired training data and manual metal segmentation, this framework provides a scalable, robust solution for enhancing diagnostic clarity in dental implant imaging. While expert validation confirms the high fidelity of the restored images, the authors emphasize that the system should function as a clinical decision-support tool, necessitating human oversight to ensure diagnostic reliability in complex or extreme clinical scenarios.