TDMAR-Net: a frequency-aware tri-domain diffusion network for CT metal artifact reduction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Metal implants cause significant artifacts in computed tomography (CT) images, impeding clinical diagnosis.
• Existing metal artifact reduction methods have limitations, including residual artifacts and reliance on paired data or multi-domain information.

Purpose of the Study:

• To propose TDMAR-Net, a diffusion model-based three-domain neural network for metal artifact reduction and CT image quality enhancement.
• To leverage priors from projection, image, and Fourier domains for improved artifact removal.

Main Methods:

• Developed TDMAR-Net, a diffusion model utilizing projection, image, and Fourier domains.
• Employed a two-stage training strategy: large-scale pretraining and masked data fine-tuning.
• Integrated a high-pass filter module in the Fourier domain and processed images in blocks to extract diffusion prior information, iteratively filling artifacts in sinogram and image domains.

Main Results:

• TDMAR-Net demonstrated superior performance compared to existing unsupervised methods.
• The method effectively removes metal-induced artifacts and enhances CT image quality.
• Validation was successful on both synthetic and clinical datasets.

Conclusions:

• TDMAR-Net effectively overcomes challenges in cross-domain information sharing for precise and robust metal artifact elimination.
• The proposed approach offers significant improvements in CT image quality in the presence of metal implants.