Multi-stage deep learning artifact reduction for parallel-beam computed tomography.

Area of Science:

• Medical Imaging
• Computational Science
• Materials Science

Background:

• Synchrotron radiation computed tomography (CT) offers high resolution but is prone to artifacts.
• Existing deep learning methods for CT artifact reduction have limitations in efficiency and effectiveness for synchrotron data.
• Artifacts can propagate through the CT computational pipeline, degrading final image quality.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for artifact reduction in synchrotron CT.
• To address limitations of current methods in handling specific artifacts and computational efficiency.
• To improve the overall quality of reconstructed synchrotron CT images.

Main Methods:

• A multi-stage deep learning framework was developed, with separate models for projection, sinogram, and reconstruction stages.
• Bypass connections were incorporated to feed raw data and intermediate outputs to subsequent stages, mitigating error propagation.
• The method was evaluated using both simulated and real-world synchrotron CT datasets.

Main Results:

• The proposed method effectively reduced artifacts across different stages of the CT pipeline.
• Evaluations demonstrated superior performance compared to existing artifact reduction techniques.
• The approach showed significant improvements in image quality for synchrotron CT data.

Conclusions:

• The novel, stage-specific deep learning method offers an effective solution for artifact reduction in synchrotron CT.
• This approach enhances image quality and overcomes limitations of previous deep learning applications in this field.
• The technique holds promise for broader adoption in scientific imaging requiring high-fidelity CT reconstruction.