Enhancing Cone-Beam Computed Tomography Image Quality Using Deep Learning-based Super-Resolution.
Authors
Affiliations (4)
Affiliations (4)
- Department of Dentistry and Oral Health, Aarhus University, Vennelyst Boulevard 9, 8000 Aarhus C, Denmark.
- Department of Conservative Dentistry, Periodontology and Digital Dentistry, LMU University Hospital, Goethestraße 70, 80336 Munich, Germany.
- Division of Endodontics, Department of Advanced Oral Sciences and Therapeutics, School of Dentistry, University of Maryland, Baltimore, 650 West Baltimore Street, Baltimore, MD 21201, USA.
- Private Practice, Centreville Endodontics, 5957 Centreville Crest Lane, Centreville, VA 20121, VA, USA.
Abstract
To develop deep learning super-resolution (SR) models for enhancing cone-beam computed tomography (CBCT). Seventy CBCT volumes (75 µm voxel size) served as references. Paired low-resolution (LR) volumes (300 µm) were simulated using physics-based degradation. Four architectures were each trained with three loss configurations (L1; L1+edge; L1+edge+perceptual): Swin Transformer for Image Restoration (SwinIR), Hybrid Attention Transformer (HAT), Hybrid Multi-Axis Aggregation Network (HMA-Net), and Dense-Residual-Connected Transformer (DRCT). Performance was assessed with peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), learned perceptual image patch similarity (LPIPS), and deep image structure and texture similarity (DISTS). Three dentists ranked the 12 configurations in 28 sets using blinded forced-choice pairwise comparisons. Correlations of metrics and expert ranks were assessed using Spearman correlation. All configurations outperformed LR input and bicubic interpolation on every metric (p < 0.05). HMA-Net trained with L1+edge+perceptual loss performed best on LPIPS (0.216) and DISTS (0.186). HAT with L1 loss achieved the highest PSNR (28.79 dB) and SSIM (0.820). Experts ranked DRCT and HAT with L1 loss, and DRCT with L1+edge loss, as the best configurations (p < 0.001 versus others). Expert rank correlated moderately only with PSNR (rho = -0.65) and SSIM (rho = -0.60; both p < 0.05). Transformer-based SR improved the perceived quality of clinical CBCT images. HAT and DRCT trained with L1 pixel loss alone, and DRCT trained with L1 and edge loss, demonstrated superior performance. Transformer-based super-resolution improved expert-perceived detail of simulated low-resolution in vivo CBCT scans without additional radiation dose, supporting further validation on directly acquired low-resolution scans.