Experimental comparison of empirical material decomposition methods for spectral CT

Area of Science:

• Medical Physics
• Biomedical Imaging
• Detector Physics

Background:

• Spectral CT using photon-counting detectors offers material decomposition capabilities.
• Non-ideal detector effects like charge-sharing and pulse-pileup introduce spectral distortions, leading to material decomposition errors.

Purpose of the Study:

• To compare the performance of a neural network estimator and a linearized maximum likelihood estimator (A-table method) for material decomposition.
• To evaluate the bias and standard deviation of material decomposition estimates using both simulated and experimental data from photon-counting x-ray detectors.

Main Methods:

• Simulations and experiments were conducted using a photon-counting x-ray detector.
• Two empirical decomposition methods, a neural network and the A-table method, were implemented and compared.
• Performance was assessed based on bias and standard deviation of material decomposition estimates for varying material thicknesses.

Main Results:

• Both neural network and A-table methods showed similar performance on simulated data.
• The neural network exhibited lower standard deviation across various material thicknesses in both collimated and uncollimated experimental data.
• Empirical methods reduced bias from 11-28% to 0.1-11% in experimental Teflon thickness studies.

Conclusions:

• Empirical methods, particularly neural networks, effectively mitigate non-ideal detector effects in spectral CT.
• Preliminary experimental feasibility of empirical material decomposition using photon-counting detectors for spectral CT is demonstrated.