Research On Low-dose CT Noise Reduction Method Based On Dictionary Learning

With the wide application of CT technology in clinical diagnosis and routine detection,its dose problem has attracted more and more attention.A large number of scientific researches have proved that excessive radiation of X-rays can cause human metabolic disorders,cause cancer and leukemia and other hazards.The reduction of the radiation dose of X-rays often degrades the quality of CT images.Using related algorithms to obtain CT images of standard quality under the condition of reducing radiation dose has become an important subject in the field of CT image applications.This thesis focuses on the problems of noise and artifacts in low-dose CT images.The main research contents are as follows:(1)A low-dose CT image noise reduction algorithm based on K-means clustering and sparse dictionary learning is proposed.The algorithm uses the K-means clustering method to divide the image clusters into multiple different classes.Both contain image blocks with similar structures;secondly,sparse dictionary learning is performed,and the number of iterations is set.When the number of iterations in sparse Bayesian preprocessing is not greater than the set number of iterations,sparse Bayesian learning is used for image blocks.Update the noise and solve the sparse coefficient;otherwise,reset the maximum number of non-zero elements to the sparsity,and then use the K-SVD algorithm to finally obtain the denoised image.By analyzing the results of different abdominal CT experiments,it can be concluded that the algorithm in this chapter can effectively suppress noise and artifacts and make the edges clear and distinct.(2)A study on low-dose CT image denoising based on nonparametric Bayesian dictionary learning is proposed.Aiming at the problem of slow convergence speed of noisy image restoration,the algorithm in this chapter is a new BPFA algorithm that combines BPFA algorithm and fast non-uniform filtering(Fast Non-Local Means,FNLM)algorithm.First,the OTSU algorithm is used to divide the target and the background;the Gaussian function smoothing parameter value is set in the FNLM algorithm,so that the noise in the target area and the background area can be well eliminated;because when the connected domain is selected,the noise in the background area is relatively low.Therefore,the non-parametric Bayesian dictionary learning is used to further denoise the background area;finally,the denoised low-dose CT image is obtained.Experiments show that the algorithm in this chapter is superior to the FNLM algorithm,the SINLM algorithm and the BPFA algorithm in terms of noise resistance,retaining more detailed information and running speed.