UPMCL-Net: Unsupervised Projection-Domain Multiview Constraint Learning for CBCT Metal Artifact Reduction.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Cone-Beam Computed Tomography (CBCT) is crucial for real-time 3D intraoperative navigation.
• Metal implants in patients cause severe artifacts in CBCT images, degrading diagnostic accuracy.
• Existing metal artifact reduction (MAR) methods fail to utilize cross-view information, leading to inaccuracies.

Purpose of the Study:

• To develop a novel unsupervised learning network for CBCT metal artifact reduction (MAR).
• To enhance image quality and diagnostic accuracy in CBCT by addressing metal artifacts.
• To improve intraoperative navigation support through artifact-free imaging.

Main Methods:

• Proposed a Unsupervised Projection-domain Multiview Constraint Learning Network (UPMCL-Net) for CBCT MAR.
• Introduced a transformer-based MultiView Consistency Module (MVCM) for cross-view projection interpolation.
• Designed a Hybrid Feature Attention Module (HFAM) for adaptive fusion of intra-view and inter-view features.

Main Results:

• UPMCL-Net effectively reduces metal artifacts in CBCT images without requiring ground truth data.
• The MVCM ensures projection-domain consistency across different views.
• HFAM adaptively integrates image features, improving artifact reduction performance.
• Experiments on real clinical data validated the network's efficacy.

Conclusions:

• UPMCL-Net offers an efficient, accurate, and reliable solution for CBCT MAR.
• The proposed method shows significant potential for improving clinical intraoperative interventions.
• This unsupervised approach overcomes limitations of current MAR algorithms by leveraging multiview information.