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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology

Background:

• Cone-beam computed tomography (CBCT) offers high spatial resolution but suffers from image quality degradation in half-detector (HD) geometry.
• Combined truncation and metal artifacts in HD CBCT are synergistic and challenging for conventional correction methods.

Purpose of the Study:

• To propose HD-TMAR, a multi-stage deep learning framework for correcting combined truncation and metal artifacts in HD CBCT.
• To systematically decompose and correct complex artifact interactions in HD CBCT imaging.

Main Methods:

• A three-stage restoration strategy: Sinogram Correction, Merging and Reconstruction, and Image Refinement.
• Utilizes sinogram-domain correction and image-domain refinement for artifact suppression and texture restoration.

Main Results:

• HD-TMAR achieved superior qualitative and quantitative results compared to existing deep learning methods for metal artifact reduction (MAR).
• Effectively suppressed severe artifacts while preserving critical dental morphology, outperforming comparative methods prone to secondary artifacts or blurring.

Conclusions:

• HD-TMAR successfully disentangles and corrects combined truncation and metal artifacts in HD CBCT.
• The framework shows significant potential for improving diagnostic performance in clinical HD CBCT applications with metallic implants.