Research On Low-Dose CT Imagimg Methods Via Deep Learning

Computed Tomography(CT),as a non-invasive medical imaging technology,can provide high-resolution tomographic images of tissues and is widely used in disease screening,diagnosis,and image-guided interventional therapy.As the number of CT scans increases,patients receiving excessive X-ray radiation during the scans increase their risk of cancer.Therefore,how to achieve low-dose imaging is an urgent problem in CT research.Low-dose CT imaging is divided into two modes presently:sparse-view and Low-mAs.Among them,the sparse-view imaging is to reduce the number of exposures of the tube per circle under a conventional scanning dose.However,the insufficient number of acquisition projections will lead to serious streak artifacts in the reconstructed image.Low-mAs imaging is to reduce the tube current or the exposure time during scanning to reduce the exposure dose of each projection.However,the reduction of the number of photons arriving at the detector causes the projection data to be contaminated by electronic noise,resulting in a large amount of noise and artifacts in the reconstructed image.In recent years,as an efficient algorithm tool,Deep Learning technology has been used as an efficient algorithm tool to solve the problems of noise and artifacts in low-dose CT imaging,and has achieved satisfactory results.The existing sparse-view image restoration networks based on deep learning are only for sparse-view image restoration of two-dimensional scanning geometry.When the input of the network is a three-dimensional spiral CT sparse-view images,the effect of the images restoration is not obvious.The existing Low-mAs image restoration networks based on deep learning are only for Low-mAs image restoration at a single noise level.When the noise level of the input images are unknown or different from the training set,the recovery result of the networks are not satisfactory.In order to realize the sparse-view image restoration of three-dimensional spiral CT and the restoration of Low-mAs image at any noise level,this paper carried out two research works based on deep learning technology:(1)For sparse-view imaging,this paper proposes a three-dimensional spiral CT sparse-view image restoration method based on multiscale wavelet residual network(MWResNet).This network combines wavelet decomposition,UNet,and residual block structure.Different from the existing network that uses two-dimensional scanning geometric sparse angle images as the training set of the network,the network in this paper uses three-dimensional spiral CT sparse angle images.By qualitatively and quantitatively comparing the restoration results of the sparse-view images of the spiral CT of the existing network,the results show that the network in this paper can effectively restore the resolution of the sparse-view images of the three-dimensional Spiral CT,suppress noise and artifacts,and maintain good image structure information.(2)For Low-mAs imaging of spiral CT,compared to image domain data,processing the projection data can obtain a more accurate noise variance estimate.Therefore,this paper proposes a blind projection recovery method based on three-dimensional wavelet residual dense network(3DWRDN)to achieve Low-mAs projections recovery of arbitrary noise level,and then reconstruct to get the denoising images.The network consists of a noise level estimation sub-network and a de-noising sub-network.The noise level estimation sub-network performs noise level estimation on three-dimensional volume data composed of multiple adjacent-angle Low-mAs projections.The denoising sub-network guides the Low-mAs projection recovery by the estimated three-dimensional noise variance map.In this paper,the three-dimensional volume projection data is used as input to effectively utilize the redundant information between adjacent angle projection data.The comparison and qualitative analysis of the recovered images with the existing image recovery networks show that the proposed network can effectively restore the Low-mAs projection of arbitrary noise level within a certain range,and reconstruct a high-resolution images with noise and artifact suppression and structure maintenance.