Study On Projection Correction And Reconstruction Image Denoising Along With Artifacts Removal Of X-ray Spectral CT

In modern medical diagnosis,X-ray computed tomography(CT)has been widely used.It uses a non-contact detection method,based on the absorption characteristics of X-rays to different materials.Attenuation characteristics acquire structural information and material composition inside an object.X-ray CT has an irreplaceable position in modern medical diagnosis.Traditional X-ray CT uses an energy integration detector to receive X-ray photons of different energy after passing through the object as a whole,reflecting the average attenuation characteristics of X-rays.It cannot distinguish X-ray photons of different energies,it is difficult to distinguish materials with similar densities,which limits the qualitative and quantitative analysis of substances.Subsequently,the X-ray spectral CT technology(Spectral CT)based on photon counting detectors appeared.The difference between energy spectral CT and traditional CT mainly lies in the use of photon counting detectors that can distinguish X-rays of different energy,which can measure objects at different absorption characteristics within range.As a result,we can improve the imaging contrast of different materials by analyzing the X-ray attenuation characteristics in different energy ranges,and achieve qualitative and quantitative measurements.We can also combine deep learning techniques to improve the imaging quality of CT.The research work of this paper was partially supported by National Key R&D Program of China(No.2016YFC0104609),National Natural Science Youth Fund(No.61401049).This thesis focused on X-ray spectral CT technology which included the principle of imaging,reconstruction algorithm,noise and artifact correction,especially improved the quality of spectral CT images and combined the deep learning method with denoising and artifacts removal in image domain.The main research contents of the thesis include:1 This paper analyzes and expounds the working principle,scanning system,imaging process,reconstruction method and other theories of X-ray CT,summarizes the characteristics and advantages of spectral CT,analyzes the causes of noise and artifacts in X-ray CT on the basis of imaging principle,and deduces the mathematical model of noise and artifacts according to the principle.This paper studies the principle of spectral CT imaging based on photon counting detectors,analyzes the reasons for the low signal noise of spectral CT images,and provides corresponding theoretical support for subsequent noise reduction and ring artifact suppression of spectral CT images.2 Aiming at the problem of poor response consistency of the detection unit of the spectral CT detector,a detection projection correction method based on the clustering algorithm is proposed.According to the working principle of the photon counting detector,the CT detection is aided by the clustering algorithm and correlation analysis.The pixels with poor consistency in the projection are located,classified and compensated.In the spectral CT detection projection,the variance in the area of the pixel with poor consistency is large.According to this,the projection is divided into regions and the corresponding regional variance is calculated.The clustering algorithm is used to classify the variance to determine the location of the pixel with poor consistency.And then use correlation analysis to classify pixels with poor consistency.As a result,targeted correction of the spectral CT detection projection can effectively improve the quality of the reconstructed image.Related experiments and results verify the feasibility of the method.3 In view of the complexity and low efficiency of the algorithm of noise reduction and artifact elimination in the image domain of spectral CT,a method of noise reduction and artifact elimination in the projection domain of spectral CT based on deep learning is proposed.Different reconstruction algorithms are used to reconstruct the spectral CT projection as training data and label data.The full convolution network and pyramid residual network are combined to form the full convolution pyramid residual network and the reconstructed image is used for training.The training model processes noise and artifacts in the reconstructed image of spectral CT at the same time,which reduces the difficulty of spectral CT correction.The experimental results show that the method based on deep learning is feasible to reduce noise and artifacts in different energy segments of spectral CT,and the performance of the proposed full convolution pyramid residual network is better than that of the common CT image noise reduction network.