A multi-stage deep learning framework for half-detector truncation and metal artifact reduction in CBCT

Cone-beam computed tomography (CBCT) is widely utilized for its high spatial resolution and compact design. However, image quality is often compromised in the half-detector (HD) geometry, where data truncation arising from the offset detector configuration coexists with metal artifacts caused by metallic implants. These artifacts interact synergistically, leading to severe image degradation that conventional correction methods fail to address.

We propose a multi-stage deep learning framework, HD-TMAR, to systematically decompose and correct combined truncation and metal artifacts in HD CBCT.

The proposed framework adopts a multi-stage restoration strategy comprising three stages: (1) Sinogram Correction, which explicitly isolates the artifact residuals from the projection data and synergizes them with structural prior-normalized features to systematically suppress global truncation biases and metal artifacts; (2) Merging and Reconstruction, which employs a specialized overlapping patching and selective replacement strategy to accurately reconstruct the truncated regions and metal traces; and (3) Image Refinement, where ImgNet further enhances the reconstructed image to restore fine anatomical textures.

Experiments using realistic simulation datasets demonstrated that the proposed method achieved the highest qualitative fidelity and quantitative metrics compared to previous deep learning MAR methods. The framework effectively suppressed severe artifacts while preserving dental morphology, whereas comparative methods suffered from secondary artifacts or blurring.

HD-TMAR successfully disentangles complex artifact interactions. By synergizing sinogram-domain correction with image-domain refinement, the framework demonstrates promising potential for clinical application in enhancing the diagnostic performance of HD CBCT systems in the presence of metallic implants.