Deep learning reduced combined truncation and metal artifacts in half-detector CBCT

KEY POINTS

• This was a simulation-based technical study using eight metal-free patient CT volumes from TCIA. Five patients were used for training, one for validation, and two for independent testing; patient-specific dental implant, crown, and bridge configurations produced 6,000 slices and 40 unique 3D artifact configurations.
• The proposed HD-TMAR framework separates the problem into three stages: sinogram-domain correction, overlapping-patch merging and FDK reconstruction, followed by an image-domain refinement network. The simulation incorporated a 120-kVp polychromatic spectrum, photon noise, and gold, titanium, and zirconium dental materials.
• In regions affected by metal, the uncorrected reconstruction achieved PSNR 25.48±3.97 dB and SSIM 0.7091±0.1268. HD-TMAR+ increased these to 35.51±1.86 dB and 0.9430±0.0188, respectively.
• The strongest conventional/deep-learning comparator in the metal-affected region, LI-ConvNet, achieved PSNR 34.46±1.39 dB and SSIM 0.9330±0.0149. Other dual-domain approaches performed less well, including DuDoNet at 33.63 dB / 0.9290 and OSCNet+ at 25.66 dB / 0.8836.
• Performance also improved outside metal-affected regions: HD-TMAR+ achieved PSNR 37.24±2.23 dB and SSIM 0.9603±0.0147, versus 33.11±2.47 dB and 0.9116±0.0267 for the corrupted reconstruction.
• Ablation experiments showed that directly feeding prior-image information could propagate distorted dental anatomy. The proposed normalized-plus-residual strategy instead reached 35.33 dB before image refinement and 35.51 dB after refinement in metal-affected regions, while better preserving tooth morphology.
• The central limitation is substantial: paired clinical half-detector datasets containing both truncation and metal artifacts were unavailable, so no real patient CBCT, registration accuracy, contouring impact, dose-calculation effect, or clinical endpoint was evaluated.

CLINICAL TAKEAWAY

SOURCE

Medical Physics