Metal artifact reduction algorithm with conditional latent diffusion model for dental cone-beam CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computational Imaging

Background:

• Metal artifacts in computed tomography (CT) present significant challenges for accurate diagnosis and treatment planning.
• Existing metal artifact reduction (MAR) techniques operate in the sinogram, projection, or image domains, with varying degrees of success.

Purpose of the Study:

• To develop an advanced metal artifact reduction (MAR) technique for dental cone-beam computed tomography (CBCT).
• To leverage a latent diffusion model (LDM) for artifact reduction and enhance the normalized metal artifact reduction (NMAR) scheme.

Main Methods:

• A latent diffusion model (LDM) was employed to generate metal artifact-reduced images, conditioned on corrupted CBCT images.
• The LDM was trained using a combined objective function of mean squared error (MSE) and learned perceptual image patch similarity (LPIPS).
• A modified normalized metal artifact reduction (NMAR) process, incorporating an automatic metal segmentation network and a secondary artifact correction network, was utilized as a prior to mitigate generative model artifacts like hallucination.

Main Results:

• The proposed LDM-enhanced NMAR method significantly outperformed classical NMAR and a state-of-the-art convolutional neural network-based MAR (CNNMAR) approach.
• Image quality improvements were quantified by reductions in RMSE (34.78 × 10⁻⁴ to 19.30 × 10⁻⁴) and increases in PSNR (49.3–54.3) and SSIM (91.2–97.2) compared to CNNMAR.
• Successful clinical implementation in dental CBCT demonstrated effective metal artifact reduction while preserving critical anatomical structures.

Conclusions:

• The developed method offers a practical and effective solution for reducing metal artifacts in dental CBCT.
• This advancement holds significant potential to improve diagnostic accuracy and treatment planning in dentistry.