Multi-Spectral Image Information Induced Spectral CT Image Reconstruction

Based on the characteristics of continuous X-ray spectrum distribution and the attenuation law of matters,spectral CT can achieve quantitative imaging of material decomposition and tissue identification,which is the representative of advanced medical CT imaging technologies.Compared with traditional CT,spectral CT can obtain spectral data in multiple energy bins according to the presetting energy thresholds,thereby generating CT images with higher tissue contrast.However,the detection of spectral CT data faces a prominent problem:the number of the detected photons in each energy bin in spectral CT is inherently less than the number of the photons detected in traditional CT,which may result in spectral CT images containing more noise and noise-induced artifacts than traditional CT images.To overcome this problem,many methods have been proposed,such as image post-processing algorithms and statistical iterative reconstruction algorithms.Image post-processing algorithms directly denoise the reconstructed images.Due to lack of accurate noise model,these algorithms may produce over-smoothed images.Statistical iterative reconstruction algorithms include a complete optical imaging model and an accurate noise distribution model,which have important significance for high-precision spectral CT image reconstruction.As the development of statistical iterative reconstruction algorithms,we propose two multi-spectral image information induced spectral CT image reconstruction algorithms,which can be summarized as:(1)The first algorithm is a non-local similarity regularized spectral CT reconstruction algorithm,which is referred to as "MECT-NSS".Since spectral CT can obtain spectral data in multiple energy bins at one scan,the reconstructed images in all energy bins are highly registered and have rich structural information.By averaging the spectral CT images in all energy bins,an average image can be obtained,which has less noise and higher resolution than the images in each energy bin.Therefore,we treat the image as a prior image and introduce it into non-local similarity regularization to reconstruct high-quality spectral CT images.Then,a comprehensive evaluation of parameter selection for the presented algorithm is conducted.We employed three sets of simulation data and a set of preclinical data to validate and evaluate the MECT-NSS reconstruction performance.Qualitative and quantitative results demonstrate that the presented MECT-NSS algorithm can successfully yield better spectral CT image quality and more accurate material estimation.(2)The second algorithm is a low-rank tensor total variation regularized spectral CT reconstruction algorithm,which is referred to as "FSTensor".The presented FSTensor algorithm is based on three considerations:1)Rich global correlation exists among the spectral CT images in all energy bins,which can be characterized by low-rank tensor decomposition.2)There is a locally piecewise smooth property in the spectral CT images,and it can be captured by a tensor total variation regularization.3)The image averaging from the spectral CT images in all energy bins are much better than,the spectral CT images with respect to noise variance and structural details and served as multi-spectral image information to improve the reconstruction performance.We then develop an alternating direction method of multipliers algorithm to numerically solve the presented FSTensor algorithm.We further utilize a genetic algorithm to tackle the parameter selection in spectral CT reconstruction.Simulation and preclinical spectral CT results demonstrate that the presented FSTensor algorithm can achieve more improvements in terms of depressing noise and preserving edge information,and acquire high-quality spectral CT images and high-accuracy basis material images.