Asymmetric decoupled dual-domain learning for CT metal artifact reduction: Integrating Fast Fourier Convolution and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Metallic implants in Computed Tomography (CT) imaging cause severe artifacts, obscuring anatomical details.
• Existing dual-domain networks for Metal Artifact Reduction (MAR) are limited by homogeneous frameworks that ignore cross-domain artifact asymmetry.

Purpose of the Study:

• To propose an Asymmetric Decoupled Dual-Domain Framework to address the limitations of current MAR methods.
• To improve the quality of CT images compromised by metallic implants.

Main Methods:

• Developed a Spectral-Geometric Inpainting Module (SGIM) using Fast Fourier Convolution for reconstructing missing projections.
• Introduced an Anatomical Texture Refinement Module (ATRM) based on Vision Transformers to mitigate residual streaks.
• Implemented a Residual-Prior Guided Fusion strategy to integrate projection space priors into image space.

Main Results:

• The proposed framework achieved superior performance compared to state-of-the-art baselines on the Synthesized DeepLesion dataset.
• Achieved a Peak Signal-to-Noise Ratio (PSNR) of 46.49 dB and Structural Similarity (SSIM) of 0.9972.
• Demonstrated superior generalization on clinical datasets and robustness against segmentation errors.

Conclusions:

• The Asymmetric Decoupled Dual-Domain Framework effectively reduces metal artifacts in CT imaging.
• The method offers high computational efficiency, making it suitable for clinical deployment.
• This approach significantly enhances diagnostic image quality in the presence of metallic implants.