Research On Low-dose CT Reconstruction Method Based On Deep Learning

CT technology,as an effective means of physical examination and disease diagnosis,has been widely used in clinic.However,the radiation produced in the process of CT examination can not be ignored.Effective measures are adopted to reduce the amount of radiation generated during CT examination,such as reducing tube voltage,tube current,exposure time,sampling Angle,etc.However,in the process of CT imaging,reducing the radiation dose will produce artifacts that do not belong to the tissue structure image,which will blur the details of the image and reduce the clarity of the image.It will even affect the diagnosis results of doctors and bring more serious harm to patients.In recent years,deep learning continues to penetrate into the field of medical imaging and has been successfully applied in low-dose CT reconstruction.Under the above background,this paper carried out research on low-dose CT reconstruction method based on deep learning technology.For low-dose reconstruction of reduced tube current and low-dose reconstruction of sparse Angle,low-dose CT reconstruction methods based on image domain multi-scale RED-CNN and projection domain interpolation convolutional neural network were proposed respectively.The specific work includes:(1)In the image domain post-processing method,the low-dose image is decomposed by NSCT(non-subsampled contour transform),and the noise information is extracted layer by layer at different scales according to the frequency segment by using its multi-scale analysis characteristics.Then,based on the decomposition coefficient images of low-dose CT,the convolutional neural network is used to extract noise information layer by layer,generate negative noise images of coefficient images,and finally obtain reconstructed CT images after image fusion.In this way,the edges and details of images are retained while noise is removed.This method selects the Red-CNN classical network model as the basic model and optimizes it,including:taking the high-frequency coefficient of low-dose CT image and the high-frequency coefficient of standard-dose CT image as the input and output of the network.The size of receptive field in the network is optimized.Based on the low-dose CT data of simulated reducer current,the simulation verification was carried out.The experimental results show that the model designed in this paper can retain the edge details of the CT image,suppress the noise caused by the low tube current very well,and achieve a good CT reconstruction effect.(2)For sparse Angle CT imaging,this paper designed the projection domain column interpolation convolutional neural network for sparse Angle low-dose CT reconstruction.Using the advantage of convolutional neural network in pixel interpolation,the projection data with missing Angle is recovered,and the reconstructed CT image is obtained by using classical filter back projection reconstruction after interpolation.The column interpolation convolutional neural network uses the high-efficiency sub-pixel convolutional neural network used in natural image super-resolution analysis task for reference,and redefines the output of the network in the way of pixel arrangement,so that it can be applied to the column interpolation task of sinusoidal image,estimate complete sinusoidal image,and complete low-dose CT reconstruction.Through the analysis of the experimental results,it can be concluded that this method can achieve good reconstruction effect at different multiples of sparse angles,remove the noise in low-dose images,and reconstruct the images more clearly.At the same time,the method uses the end-to-end method to estimate the sinusographs,which has obvious speed advantages compared with the traditional methods.