A neural network-based method for spectral distortion correction in photon counting x-ray CT.

Area of Science:

• Medical Imaging
• Materials Science
• Artificial Intelligence

Background:

• Spectral CT utilizes energy-dependent X-ray attenuation for material composition analysis.
• Photon counting x-ray detectors (PCXDs) enable spectral CT, particularly for K-edge contrast agents in soft tissue imaging.
• PCXD measurements suffer from spectral distortions and noise, limiting accuracy.

Purpose of the Study:

• To develop and validate a novel artificial neural network (ANN)-based spectral distortion correction method for PCXD spectral CT.
• To improve material decomposition accuracy and K-edge contrast agent imaging.
• To integrate denoising techniques for enhanced robustness.

Main Methods:

• Developed a micro-CT system with a PCXD capable of 4 energy bins and full-spectrum acquisition.
• Modeled spectral distortions using experimental data from radioactive sources.
• Trained an ANN to learn and correct spectral distortions in projection data.
• Applied joint bilateral filtration for post-reconstruction denoising.
• Validated the method using simulations and experimental data from phantoms and a mouse model.

Main Results:

• ANN-based distortion correction significantly improved material decomposition accuracy, reducing measurement error by up to 50% in simulations and 41% in experimental full-spectrum data.
• Full-spectrum acquisition with correction outperformed 4-energy bin mode.
• Iodine detectability improved to 4-6 mg/mL with distortion correction.
• Denoising further enhanced ANN training robustness and decomposition accuracy.

Conclusions:

• ANN-based spectral distortion correction is experimentally feasible and highly effective for PCXD spectral CT.
• The proposed method enhances material decomposition accuracy and K-edge imaging capabilities.
• The computational efficiency allows for integration into existing CT reconstruction pipelines.