Research On Noise Reduction And Material Discrimination Of X-ray Spectral CT Images Based On Photon Count Detector

As is well-know,in the large-scale outbreak of COVID-19 worldwide in the spring of 2020,the clinical diagnostic technique of medical C T is an important way to evaluate the pulmonary inflammation of patients with COVID-19.X-ray spectral CT originated from medical CT technology.Different from traditional medical CT which uses integral detector to obtain the whole information of the patient's body tissues,it can use photon count detector to obtain the X-ray attenuation characteristics of the body tissues in different energy intervals.Based on this difference,spectral CT can not only improve the image contrast of materials with similar attenuation coefficients,but also make qualitative and quantitative analysis of the scanned objects,which has attracted worldwide attention.The photon count detector can detect the photons(visible light)after the interaction of X-ray and material with different energies,and then obtain spectral CT reconstruction images with different energy intervals to identify and analyze different materials.However,the number of photons detected by X-ray spectral CT in a specific energy range or interval is limited,resulting in low SNR characteristics of the projected data and high image noise.Therefore,it is still an urgent problem to carry out efficient spectral CT image noise reduction and high-precision material analysis and discrimination.For this reason,based on the ministry of science and technology research and development of the national key project(project number: 2016YFC0104609),national natural science fund project(project number: 61401049),basic science and research in cutting-edge technologies in chongqing special project(project number:CSTC2016JCYJA0473),the basic scientific research funds in universities were project(project number: 2019CDYGYB019)and chongqing education commission science and technology research projects(project number:KJQN201904007),we studied the methods of X-ray spectral CT image noise reduction and material recognition based on deep learning.The research contents of this paper mainly include:(1)On the basis of extensive investigation of a large number of related literatures in domestic and overseas,we first expounded the research background and significance,and summarized the research progress of X-ray spectral CT technology in domestic and oversea.Then,we analyzed the research status,developme nt trend and key scientific problems of X-ray spectral CT image denoising technology and material discrimination technology.Finally,we clarify the main content of this paper.(2)This paper analyzes the basic theory of X-ray spectral CT technique.In view of the characteristics of X-ray energy spectral CT imaging technology,we do some research on the material discrimination of complex measured objects.Combining with the deep learning related algorithm,we constructed the spectral CT image denoising model and spectral CT material discrimination model.(3)Research on feature extraction model of X-ray spectral CT image based on deep learning.Although the X-ray spectral CT system built on the photon count detector can detect the attenuation characteristics of X-ray with different energies,the number of X-ray photons detected in a narrow energy range is small,resulting in a low SNR of reconstructed spectral CT images.Therefore,in order to improve the feature extraction unit of deep learning,this paper studies a new feature extraction structure----pyramid residual module,so as to effectively extract feature information of energy spectral CT images.Different from VGGNet and ResNet,pyramid residual module can effectively reduce the parameters of network.It can not only accelerate the convergence speed of network,but also reduce the risk of network overfitting and gradient explosion,so it has better feature extraction ability.This lays a foundation for further research on deep learn-based image denoising and material analysis.(4)Study on noise reduction technique of X-ray spectral CT image based on pyramid residual network.Aiming at the serious problem that the limited number of X-ray photons detected in a narrow energy range leads to the strong noise of spectral CT reconstructed images,this paper constructs a noise model and creates a data set according to the actual spectral CT data characteristics.Based on the image processing capability of pyramid residual module,we designed a noise reduction network.Compared with DnCNN and REDCNN,the proposed network in this paper obtained better results,and its high-quality image data sources provided high-quality spectral CT images for subsequent material discrimination studies.(5)Study on X-ray spectral CT material discrimination based on fully convolutional pyramid residual network.The traditional base material discrimination method of X-ray spectral CT is to reconstruct the reconstructed images.This method is easily affected by noise in the process of multi-material reconstruction,so it is difficult to obtain excellent material discrimination accuracy.Therefore,we designed a fully convolutional neural network based on the pyramid residual module and studied the spectral CT material discrimination technology.By introducing the structure of encoder-decoder and skip connections,the algorithm can detect the different regions of materials in the spectral CT image and achieve the purpose of material recognition.The proposed network feature dimension can effect ively reduce parameters by slowly increasing the network feature dimension and improve the utilization efficiency of feature map in shallow layer network by conducting feature map with skip connections.In the mouse spectral CT image data set,we verified the method.Compared with other deep learning methods such as SegNet,FCN-8s and U-Net,the proposed algorithm obtained better segmentation results.At the same time,compared with the traditional base material method,the proposed method has higher material discrimination accuracy.