Research On Key Technologies For Image Processing And Analysis Of X-ray Medical Computed Tomography Imaging

Owing to fast imaging interior tissues and organs of the patient body without invading it,X-ray medical computed tomography(CT)imaging has widely applied to clinical diagnosis and therapy.However,ionizing radiation of CT scanning will accumulate in the patients’ body and induce certain radiation damages.Hence,it demands the X-ray is low dose during CT scanning.Low dose can achieve by reducing operating voltage and current of the X-ray tube or reducing exposure time,but X-ray photons collected by the detector will decrease,inducing Poisson noise.This will certainly deteriorate signal-to-noise ratio of the collected projection data.Then the reconstructed CT image will be interfered by noise and artifacts.Therefore,in the case of low dose X-ray radiation,it has important theoretical significance that how to ensure the quality of medical CT imaging and ensure its clinical diagnostic value,and how to conduct the researches on the key technologies of low dose CT imaging denoising.Particulaly for Coronavirus disease 2019(COVID-19)lesions,chest CT imaging diagnosis is more intuitive and faster than biometric diagnostic techniques,such as reverse transcription-polymerase chain reaction(RT-PCR).Hence,it is of importance to perform segmenting COVID-19 lesions from medical CT images.Based on the information process of X-ray medical CT imaging and CT image analysis and processing,this thesis respectively carried out the in-depth researches on low dose CT projection data denoising,low dose CT image post-processing denoising,and COVID-19 lesion segmentation from chest CT images.These researches act as the points to represent the key techniques for image analysis and processing for X-ray medical CT imaging.The researches of the thesis mainly include:(1)The fundamental theory and technical methods of X-ray medical CT were studied fully and systematically,and the developing process of X-ray CT imaging technology and its research status of home and abroad were clarified.The challenges of low dose CT reconstruction and image quality degradation were described clearly.That is how to decrease the X-ray dose as much as possible for reducing risks to the patients,while guaranteeing diagnostic performance of the reconstructed CT images.The key techniques of CT image analysis and processing were described clearly,and the challenge of segmenting COVID-19 lesions from chest CT image and its corrected segmentation were elaborated.On the basis mentioned above,the ideas of research routes of low dose CT projection data denoising,low dose CT image post-processing denoising,and COVID-19 lesion segmentation from chest CT images were clearly illustrated.(2)An algorithm for low dose CT projection data denoising was investigated.Existed low dose denoising methods directly perform on low dose CT images readily losing subtle structures and details,while raw projection data processed by convolutional neural network(CNN)is scarcely reported.To deal with the problem,attention guided residual dense CNN(Att RDN)was presented for low dose projection data denoising.First,aided by the attention mechanism,Att RDN fully utilized advantages of feature fusion and global residual learning to extract noise from polluted low dose CT projection data;then,the denoised projection data obtained by subtracting the extracted noise from input low dose projection data;finally,the CT image was reconstructed from the denoised projection data by analytic reconstruction.The experimental results demonstrate that the proposed Att RDN has better denoised performance qualitatively and quantitatively than the compared methods in our study.(3)The potential risks of X-ray radiation of CT scanning to patients have been double focused on by the public,and simply reducing radiation dose will result in quality degradation of the reconstructed CT image,and affect the diagnostic performance.To address this issue,a noise-learning low dose CT image post-processing algorithm referred as to hybrid loss least squared generative adversarial network(H-LSGAN)was proposed,which combined the least square loss,structural similarity loss and L1 loss.Specifically,complex noise distributed in the low dose CT image was extracted by the generator.Then the complex noise was subtracted from the input to obtain the denoised CT image.The least squared loss can penalize the denoised CT image,which discriminated to be normal dose but far away from normal dose CT data distribution,and pull it toward the decision plane;structural similarity and L1 losses are utilized to preserve the image texture and sharpness.The experiments show that the presented H-LSGAN can remove noise and suppress artifacts distributed in the low dose CT image.The textural statistical analysis further confirmed that the restored CT images closed to the corresponding normal dose CT images.(4)Due to infection area with high variation,low contrast to surrounding normal tissues,and blurred of lesion boundary,segmenting COVID-19 lesions from chest CT image is a very challenge.Moreover,the available CT dataset is limited that greatly hinder machine-learning techniques applied for the fight against COVID-19.To solve the problem,a deep learning-based method of segmentation of COVID-19 lesions from chest CT images was presented,which was known as pyramid pooling module improved UNet(PPM-UNet).This method firstly adopted pyramid pooling module to replace the original skip-connections in the UNet architecture,and then utilized the global attention mechanism to enhance the representation capability of the neural network.First,a coarse model was obtained by pre-training the PPM-UNet on a CT dataset containing 1600 pseudo labels of COVID-19.Then the coarse model was fine-tuned on the standard COVID-19 CT dataset including 100 labels to produce a fine PPM-UNet for COVID-19 lesion segmentation.The experimental results show that the performance of PPM-UNet exceeding the other state-of-the-art segmentation methods.This thesis researched on the key technologies for low dose CT projection data denoising,low dose CT image denoising,and COVID-19 lesion segmentation from CT images.These works based on image processing and analysis of X-ray medical CT imaging.And all of these involve the fields or disciplines of nuclear radiation dose,X-ray detection principle,optical imaging,biomedical engineering,computed image processing and reconstruction,which integrating the crystallization of research works of various disciplines and practicing the researches of cross-disciplinary.At present,the low dose X-ray medical CT imaging is both a hot issue in science and technology,and rapid development at the forefronts.Studies of this thesis not only own broad prospects in clinical application,and effectively promote the indispensable CT imaging for accurate medical technology in our country.