Fully connected neural network for virtual monochromatic imaging in spectral computed tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Photon Counting Detectors

Background:

• Spectral computed tomography (SCT) enables material decomposition and quantitative reconstruction.
• Nonideal detector effects (charge sharing, pulse pileup, escape) cause spectral distortion, hindering accurate material decomposition.
• Precise system modeling for practical SCT is challenging due to these physical effects.

Purpose of the Study:

• To develop a data-driven spectral information extraction method for SCT.
• To generate virtual monochromatic attenuation maps without prior spectral information.
• To improve material decomposition accuracy and image quality in SCT.

Main Methods:

• Utilized a fully connected neural network for spectral information extraction.
• Trained the neural network using data from various substances.
• The network directly outputs virtual monochromatic linear attenuation coefficients.

Main Results:

• Achieved accurate virtual monochromatic linear attenuation coefficients.
• Demonstrated outstanding performance in denoising and artifact suppression.
• The method shows good generalization ability across different SCT system settings and substances.
• Achieved low testing mean square errors (MSE).

Conclusions:

• The proposed neural network method effectively extracts spectral information for SCT.
• This approach bypasses the need for complex physical system modeling.
• The method enhances material recognition and image quality, offering a versatile solution for SCT systems.