A Study On Denoising For Low-dose CT Imaging

Computerized Tomography(CT),one of the most commonly used medical imaging devices,plays an important role in clinical diagnosis and treatment,but high doses of X-ray ionizing radiation can cause damage to human tissues.In general,reducing tube current,peak voltage,and shorten the slice scan time are three effective ways to reduce CT radiation dose.The reduction of the tube current will decrease the number of collected X-ray photons,and the projection data will be contaminated by noise,further causing for a large amount of noise and artifacts in the reconstructed image.Reducing the peak voltage causes the X-ray penetration to weaken,which will give rise to appear hardening artifacts.Short scan or sparse scan can shorten the slice scan time,but it also brings the problem of incomplete image reconstruction data,resulting in streak artifacts.How to improve the quality of CT imaging under low-dose has become an important research topic.At present,there are three type solutions for the above problems that can balance the reduction in radiation dose and the improvement in image quality.One type is the projection domain denoising method,also known as the preprocessing method,which mainly applies the appropriate denoising method to improve the quality of the original projection data,increase the signal-to-noise ratio(SNR)of the image,and then the filtered back projection(FBP)method is used to reconstruct the final image.But in the case of lower doses,the efficiency of such methods is reduced and the denoising effect is becoming poorer.The other type is the image domain denoising method,which mainly performs denoising processing on the reconstructed CT image,also known as post-processing method.Although the image quality can be improved,the loss of the image details is caused by not using the noise statistical characteristics of the projection data.The last category is the statistical iterative reconstruction(SIR).Such methods based on the statistical characteristics of the projection data can develop an optimal objective function based on maximum likelihood or penalty likelihood,which can accurately predict the physical process of low-dose CT imaging.The expression variable is directly a pixel or voxel in the CT image domain,so it can reconstruct a higher quality CT image from the projection data containing noise,and is suitable for image reconstruction under the condition of incomplete data,reducing the streak artifacts,but such methods have slow convergence and it needs to be equipped with high-performance computers for iterative reconstruction.When the iterative reconstruction intensity is high,the image will appear excessively smooth,which makes the image texture features of CT reconstruction appear different.This paper focuses on the denoising of low-dose CT from both the projection domain and the image domain,mainly including the following aspects:(1)Low-dose CT image denoising based on non-penalty function and wavelet-TGV in projection domain.Aiming at the noise statistical characteristics of low-dose CT projection data,the shortcomings of wavelet threshold processing,total variation(TV)minimization and wavelet-TV in the denoising of projection domain are analyzed.The denoising model based on non-convex penalty function and wavelet-TGV is established.The model utilizes the high-order derivative information of total generalize variation(TGV),which can eliminate the ladder phenomenon and block artifacts brought by TV minimization denoising process,and make the objective function more sparser than the wavelet-TV method under the constraint of non-convex penalty function.Using the split augmented Lagrangian shrinkage algorithm(SALSA)for optimization,the results of denoising and reconstructing are compared with other methods from visual effects and quantitative analysis.It is proved that the proposed method can improve the image quality while maintaining the boundary and texture details.(2)Joint denoising method based on RNLM.For the optimization solution in(1),a large number of iterative operations are required,and will have to led an inefficient solution.The projection data is converted using variables stabilization transformation(VST)from the unsteady state of the signal to the steady state,and then the conventional image denoising method such as block matching three-dimensional cooperative filtering(BM3D)can be applied to improve the efficiency of denoising in the projection domain.In order to further improve the image quality,the reference-based non-local means(RNLM)filtering method is combined with the denoising of the projection domain and the image domain.The weighted balance processing is derived from the image denoised in the projection domain and the image domain respectively.The denoising effect of the method is remarkable verified by experiments.The advantages of both the projection domain denoising and the image domain denoising can be fully utilized.It is especially suitable for cone beam CT using FDK(Feldkamp-Davis-Kress)reconstruction method.(3)Low-dose CT denoising based on shearlet transform and autoencoder.Since the first two methods need to access the original projection data that is opaque to the most of users,the practical application is limited to some extent.The shearlet can better represent the texture features and multi-scale image decomposition characteristics in the high frequency space.A convolutional neural network based on shearlet transform and denoising autoencoder is designed by using residual network learning method for image denoising of low-dose CT.The denoising method simplifies the denoising problem model of low-dose CT images.Without accessing the projection data,high-dimensional features of the image can be captured from a large amount of existing training data,and experimental verification is performed through simulation data and clinical data.The method makes full use of the existing large amount of CT image resources.When the tube current is reduced by 75%,the method enhances noise suppression and edge structure information retention,thereby improving image quality and ensuring detection capability of small lesion areas.