Artifact Correction Of CT/CBCT Image Based On Deep Convolutional Network

Nowadays,X-ray computed tomography(CT)is widely used to help doctors diagnose diseases.However,due to the health risk of radiation exposure to patients,researchers need to seek ways to reduce the radiation dose.The most widely used method is lowering the tube mAs.Nevertheless,low-dose techniques cause streaking artifacts and would degrade image quality,which affects clinical diagnosis significantly How to reduce the dose of CT image while ensuring the quality of image has become one of the research hotspots of CT imaging.To cope with the problems associated with low-dose CT,a number of algorithms have been designed;these algorithms can be roughly divided into three categories,namely,sinogram filtering techniques,iterative reconstruction,and postprocessing methods.These algorithms often need to estimate and assume a specific noise model,but the artifact of low dose CT is difficult to determine its noise model,which leads to inaccurate artifact estimation.However,deep learning can make full use of the characteristics of the data,which can make the prediction of low-dose CT more accurate,and it has become another hot spot of low-dose CT artifact correctionCone beam CT(CBCT)was also designed for the low dose of radiography Compared with the traditional CT,it has the characteristics of low dose,high X-ray utilization rate,fast data acquisition and so on.However,because of its physical and technical problems of hardware,CBCT images are prone to serious scattering artifacts,which limits its wide clinical application.Therefore,the problem of artifact correction of CBCT images has become an urgent problem to be solved in the industry.To cope with the scatter artifact in CBCT,a number of algorithms have been presented,mainly divided into two categories,hardware processing methods and software processing methods.Among them,the traditional Monte Carlo simulation method is the gold standard for scattering projection estimation,but its clinical application is greatly limited by its excessive photonic transport simulation time.Therefore,it is still difficult to design an accurate and fast CBCT artifact correction algorithm.In this paper,the imaging basis of CT/CBCT is reviewed systematically,and the basic content of convolution neural network is learned.Then,aiming at the above two image artifacts,the following correction methods based on deep convolution neural network are proposed respectivelyFirstly,in order to solve the problem of artifact correction of low dose CT images after reconstruction,a noise suppression algorithm TLRCNN based on transfer learning and residual network is proposed.In this method,a convolution neural network composed of residual blocks is constructed to learn the mapping from LDCT image to noise map.Firstly,the noisy natural image dataset is used to pre-train the network,and then the LDCT dataset is used to fine-tune the model.Finally,the corrected CT image can be obtained by subtracted the noise map predicted by the trained model from the LDCT image.The experimental results of simulation data and clinical data show that compared with commonly used denoising algorithms,TLRCNN can restore LDCT images effectively and with high quality,and its image quality can be improved qualitatively and quantitativelySecondly,in order to solve the problem of extensive calculation time of CBCT scattering artifact correction,a fast scatter correction algorithm FSCCNN based on gMMC Monte Carlo simulation and U-Net is proposed.This method mainly operates in the projection domain.Firstly,the FDK algorithm is used to reconstruct the original CBCT projection,and then the rigid registration is carried out with the planning CT reconstruction.Then the acceleration strategy is used to simulate the sparse angle scattering projection on the gMMC system.And a trained U-Net network model will be used to restore the full angle scattering projection.Finally,the scattering projection is removed from the original projection,and the corrected CBCT image can be obtained by FDK.The experimental results show that compared with the planned CT image,the image quality corrected by this algorithm is obviously improved,and so is the accuracy of CT value.All the evaluation indexes are very close to the planning CT.The most important point is that the method can complete the CBCT scatter correction within 10 s,and its efficiency is greatly improved compared with the traditional MC simulation.