Structure-Preserving Two-Stage Diffusion Model for CBCT Metal Artifact Reduction.

Area of Science:

• Dental Imaging
• Medical Image Analysis
• Artificial Intelligence in Healthcare

Background:

• Metal artifacts in Cone-Beam Computed Tomography (CBCT) significantly hinder accurate dental diagnosis.
• Existing deep learning methods for metal artifact reduction (MAR) often fail to preserve surrounding anatomical structures and generalize poorly to real-world data due to domain gaps.

Purpose of the Study:

• To develop a robust MAR method for CBCT that emphasizes structure preservation and domain generalization.
• To overcome the limitations of current deep learning approaches in handling real-world dental imaging data with metal implants.

Main Methods:

• A two-stage diffusion framework was proposed: Stage I uses a structure-aware diffusion model trained with fused intraoral scan (IOS) and CBCT data for edge map extraction.
• Stage II employs these edge maps as structural priors for MAR, incorporating a segmentation-guided sampling (SGS) strategy to enhance structure preservation during inference.

Main Results:

• The proposed method demonstrated superior artifact reduction capabilities on both simulated and real-world CBCT datasets.
• Significantly better preservation of critical dental anatomical structures around metal implants was achieved compared to existing methods.

Conclusions:

• The two-stage diffusion framework with structure-aware edge map extraction and SGS effectively addresses MAR challenges in CBCT.
• This approach offers improved diagnostic accuracy in dental applications by enhancing both artifact reduction and structural integrity.