Deep learning-based noise reduction for ultra-low-dose dental CBCT images using paired datasets from different domains

This study presents a novel approach to ultra-low dose (ULD) dental cone beam computed tomography (CBCT) noise reduction using deep-learning. We address the challenges of reducing noise in ULD images while preserving anatomical structures for clinical application. For supervised learning, we constructed two datasets: one from scanned head phantoms and another from simulated clinical data. The proposed method pre-trains a U-Net-based model on paired phantom dataset and fine-tunes it with simulated dataset, balancing noise reduction in hard tissues and preservation of soft tissue structures. In qualitative evaluation, a model trained on phantom dataset achieves notable noise reduction but tends to over-smooth soft tissues. Conversely, a model trained on simulated dataset better preserves anatomical structures but shows less effective noise reduction. The proposed method, which combines the strengths of both approaches, successfully maintains soft tissue details while significantly reducing noise. Quantitative analysis shows consistent performance across both phantom and simulated dataset. Relative to the original ULD images, the proposed method achieves peak signal-to-noise ratio (PSNR) improvements of 3.102 dB and 2.706 dB, and structural similarity index measure (SSIM) increases of 8.3% and 6.3% for phantom and simulated dataset, respectively. The proposed method shows promise for clinical implementation of deep-learning-based noise reduction. This approach is expected to provide diagnostic-quality CBCT scans at extremely low radiation exposure.