Study On Key Techniques Of Biomedical Material Decomposition Based On X-ray Spectral CT

With the rapid development of information technology such as artificial intelligence,human beings have entered the era of smart medical care.Machine learning algorithms represented by deep learning have also begun to be widely used in various fields of medical imaging.X-ray computed tomography(X-CT),as one of the important components of smart medical care,plays an important role in medical clinical diagnosis.It utilizes the difference in X-rays attenuation of different materials,and can present its internal structure and details through the tomographic image without destroying the object itself.However,the traditional X-CT adopts integral detection mode to acquire X-ray photon information,which will cause the loss of attenuation characteristics to some extent,and it is difficult to meet the demand for precise diagnosis and treatment.Thus,X-ray spectral CT imaging technology came into being.X-ray spectral CT imaging benefits from the emergence and development of X-ray energy-resolved photon-counting detection technology.It uses the photon-counting detector to obtain the X-ray attenuation characteristics of detected objects in different energy ranges.According to these differences,not only can the imaging contrast of materials with similar attenuation coefficients be improved,but also the qualitative and quantitative analysis of the scanned objects can be performed.Among them,the material decomposition based on X-ray spectral characteristics has become an important research direction of spectral CT imaging technology.The research work of this paper relied on the National Key Research and Development Program of the Ministry of Science and Technology(No.2016YFC0104609),the National Natural Science Foundation of China(No.61401049),and the Chongqing Basic Science and Frontier Technology Research Special Project(No.cstc2016jcyjA0473).This thesis focused on X-ray spectral CT technology,and studied the biomedical material decomposition,which not only improved the traditional projection decomposition but also combined the deep learning method with material decomposition in image domain.The main research contents of the thesis include:(1)Conducting in-depth study of X-CT imaging principles and systems,reconstruction methods and technology,detailing the basic theories and models for spectral CT material decomposition.Aiming at the noticeable noises in projection data of X-ray spectral CT,the Split-Bregman algorithm based on compressed sensing theory was selected for image reconstruction.It is the important prerequisite for spectral CT material decomposition.(2)Aiming at the problem of complex projection integral calculation in the material decomposition process based on pre-reconstruction,a dual-energy CT projection decomposition method based on equivalent monochromatic energy was studied.By calculating the average attenuation coefficients of a specific material region from the dual-energy reconstructed images,and obtaining the equivalent monochromatic energy corresponding to the two energy spectra according to the attenuation curve fitted by the standard X-ray attenuation database provided by NIST,the projection decomposition is simplified to a linear problem.Simulation and experiments have well verified the availability of the proposed method.(3)Aiming at the problem that material decomposition based on post-reconstruction is very susceptible to the influence of actual data noise and cannot achieve the ideal decomposition effect,a deep learning-based spectral CT multi-material decomposition method was studied.Due to the relatively small amount of spectral CT image data and the characteristics of single-channel grayscale images,FC-DenseNets was selected as the training network.The data set was constructed using the reconstructed spectral CT images.The experimental results showed that the deep learning method has great advantages over the traditional method of basis material decomposition,which cannot effectively distinguish the components with similar attenuation and cope with the high-noise situations.FC-DenseNets not only accurately identified and decomposed the bone,lung and soft tissue regions of mouse,but also achieved ideal decomposition results in the case of large image noise.