Correction of x-ray absorption spectrum distortion using adversarial neural networks

Abstract

Energy spectrum CT can provide higher CT imaging resolution and thus play an important role in clinical diagnosis. X-ray Absorption Spectrum (XAS) instrument based on photon counting detector has great potentials in energy spectrum CT. The main problems of PCDs are “Stacking effect” and “charge sharing”, which can cause spectrum distortion and therefore depress imaging quality. It is quite complicated to correct spectral distortion through hardware improvements and physical simulation. Hence, we proposed a method for x-ray absorption spectrum data correction based on adversarial neural networks. The generator of our proposed can generate corrected spectrum after training. Spectrum data for correction in our experiments was collected by laboratory equipment and the ground-truth was obtained by simulation software. Real data and simulated data were feed to the network to train the generator and discriminator simultaneously. Results of our experiment illustrated that the well-trained network can effectively recover the spectrum from the distorted spectrum data. We also evaluate the imaging quality with the corrected spectrum data. It can be seen that the quality of CT reconstruction image can be significantly enhanced through corrected spectrum data. Our method of using adversarial neural networks to generate noise-free x-ray spectrum provides new ideas for the clinical application of energy spectrum CT.