Deep-learning-based ring artifact correction for tomographic reconstruction.

Area of Science:

• Medical Imaging
• Computational Science

Background:

• X-ray tomography provides high-resolution, non-destructive 3D structural observation across various research fields.
• Ring artifacts, caused by detector nonlinearity, degrade tomographic reconstruction quality and introduce bias.

Purpose of the Study:

• To develop an advanced ring artifact correction method for X-ray tomography.
• To improve image quality and reduce bias in tomographic reconstructions.

Main Methods:

• A novel ring artifact correction method utilizing a residual neural network (ResNet).
• The network leverages wavelet coefficients and residual blocks for high-precision artifact removal with low computational cost.
• A regularization term is employed for accurate stripe artifact extraction in sinograms.

Main Results:

• The proposed ResNet-based method effectively suppresses ring artifacts in both simulated and experimental X-ray tomography data.
• The technique demonstrates superior preservation of image details compared to existing methods.
• Transfer learning enhances the ResNet model's robustness, versatility, and computational efficiency, addressing limited training data.

Conclusions:

• The developed ResNet method offers a robust and efficient solution for ring artifact correction in X-ray tomography.
• This approach significantly improves image quality, making X-ray tomography more reliable for detailed 3D structural analysis.
• The use of transfer learning ensures practical applicability even with limited datasets.