Polychromatic neural CBCT reconstruction through density-attenuation modeling

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Monochromatic cone beam computed tomography (CBCT) algorithms are prevalent but suffer from beam-hardening artifacts due to energy-integrating detectors.
• These artifacts arise because current detectors cannot resolve photon energy levels, leading to inaccurate attenuation measurements.

Purpose of the Study:

• To develop a novel polychromatic 3D reconstruction technique for CBCT.
• To mitigate beam-hardening artifacts without requiring additional prior information.

Main Methods:

• A coordinate-based neural representation was employed for polychromatic 3D reconstruction.
• The method models attenuation at a reference energy level (E0) and its derivative.
• A neural network was implemented to optimize intermediate density values and a composite attenuation function, learning a monotonic relationship between density and attenuation.

Main Results:

• The proposed technique significantly improved reconstruction quality in various scenarios.
• Quantitative improvements were demonstrated using a synthetic numerical phantom.
• Enhanced visual quality was observed in real-world clinical examples.

Conclusions:

• The developed polychromatic reconstruction method effectively reduces beam-hardening artifacts in CBCT.
• This neural network-based approach offers a promising solution for improving CBCT image accuracy and quality.