CT sinogram-consistency learning for metal-induced beam hardening correction.

Area of Science:

• Medical Imaging
• Computerized Tomography
• Artificial Intelligence

Background:

• Beam hardening artifacts in polychromatic CT arise from highly attenuating materials, causing sinogram inconsistency and reconstruction errors.
• These artifacts manifest as streaking and shading, degrading diagnostic image quality.

Purpose of the Study:

• To propose and validate a sinogram-consistency learning method for reducing beam hardening artifacts in CT.
• To address the challenge of inconsistent sinograms caused by metal implants in CT scans.

Main Methods:

• A deep learning approach is employed to repair inconsistent sinograms by removing metal-induced beam hardening factors.
• The method utilizes simulated training data and patient implant type-specific models to simplify the learning process.
• Sinogram consistency is enforced by learning to remove primary beam hardening effects along metal traces.

Main Results:

• The proposed method was validated on CT scans of pelvises with simulated hip prostheses.
• The deep learning model successfully corrected sinogram inconsistency by selectively extracting beam hardening sources.
• The method demonstrated effectiveness even with different anatomical areas between training and testing data.

Conclusions:

• This work presents the first deep learning-based sinogram correction method for beam hardening reduction in CT.
• Unlike conventional regularization methods, this deep learning approach offers potential for broader application across various CT images.
• The sinogram-consistency learning method shows promise for improving CT image quality in the presence of metal artifacts.