A Denoising Diffusion Probabilistic Model for Metal Artifact Reduction in CT.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Metal artifacts corrupt CT images, hindering diagnosis.
• AI, including CNNs and GANs, has shown promise for metal artifact reduction (MAR).
• Denoising diffusion probabilistic models (DDPMs) offer advanced image generation capabilities.

Purpose of the Study:

• To propose and evaluate a DDPM-based approach for inpainting missing sinogram data for improved MAR.
• To assess the generalization capabilities of an unconditionally trained DDPM for diverse metal implants.
• To compare the DDPM-MAR performance against NMAR, CNN-based, and GAN-based methods.

Main Methods:

• An unconditionally trained DDPM was developed for sinogram data inpainting.
• The DDPM-MAR technique was evaluated using SSIM and PSNR metrics.
• Performance was compared against NMAR, CNN-MAR, and GAN-MAR approaches.
• Clinical CT images with simulated metal artifacts were used for further evaluation.

Main Results:

• The DDPM-based approach significantly outperformed NMAR, CNN-based, and GAN-based MAR methods in SSIM and PSNR.
• Clinical evaluation demonstrated superior image quality with the DDPM-MAR technique.
• AI-based methods generally showed better MAR performance than non-AI NMAR.

Conclusions:

• The proposed DDPM-based MAR technique effectively reduces metal artifacts in CT images.
• This AI-driven method enhances image quality and holds potential for improving diagnostic accuracy.
• Unconditional training of DDPMs shows promise for robust MAR across different metal implants.