Research On Low Dose X-ray Fluoroscopy Image Denoising

X-ray has been widely used in medical imaging diagnosis because of its strong penetrating ability,which greatly improves the accuracy of the doctor’s diagnosis of the lesion.However,X-ray irradiation to the body at the same time,the body also has some damage,long-term high doses of X-ray radiation can greatly increase the prevalence of cancer,so reduce the X-ray radiation dose can effectively reduce the X-Ray damage to the human body.At present,many research groups at home and abroad are trying to study ways to improve the quality of low-dose perspective images.But most of the noise reduction research are concentrated in the CT image,but also made a lot of results.There are few studies on noise reduction of low dose X-ray fluoroscopy images.The experimental data processed in this paper is a series of low-dose X-ray fluoroscopic images that can be rotated without the need for a normal dose X-ray technique if the results can be produced in low-dose X-ray fluoroscopy image noise reduction studies Equipment,but only to reduce the amount of current into the tube,which greatly reduces the cost of equipment procurement.Therefore,it is of great practical value to reduce noise in severe low-dose X-ray fluoroscopy images.The main work of this paper is to analyze the noise in the low-dose X-ray perspective image,and to explore and improve some noise reduction algorithms.In this paper,we first study the noise characteristics in the experimental data.By analyzing the image of different thickness plexiglass,it is concluded that the noise in the low-dose X-ray fluoroscopic image is the multiplicative noise which is similar to the Gaussian distribution.Secondly,based on the noise characteristics,the noise reduction algorithm based on statistics and learning is explored:principal component analysis(PCA)and K-mean singular value decomposition(K-SVD).The principal component analysis algorithm extracts the feature of the image data,and in this process,most of the noise in the image is filtered out,but the algorithm in the process of suppressing the noise,some of the details of the image information is also filtered.Based on the dictionary learning algorithm,through the continuous iteration of the dictionary to get a better representation of the image,but because there is no noise without the original image of the dictionary training,this article uses a noisy image for training,so the experimental results Not very good.In addition,the nonlocal mean algorithm(NLM)is explored,which uses the redundant information of the image to perform noise reduction.In this paper,the process of eliminating additive noise based on LPGPCA algorithm is introduced.The algorithm introduces the concept of local grouping.The block is used as the basic unit of image processing.The processing of each pixel is the pixel block centered on the pixel to be processed.Locate other similar image blocks in the local image block and use the results to find the training set.In the improved algorithm,the selection of the training sample set is extended to the similarity between the adjacent images according to the experimental data(100 continuous low dose X-ray fluoroscopy image,there is a certain rotation angle between adjacent images)In all similar images,the final set of training samples consists of similar pixel blocks in all similar image sets in the local image block.The expanded set of training samples greatly increases the possibility of proposing the main feature components,and the time complexity of the algorithm is controlled within a certain range.The improved algorithm is divided into two parts,the first part is the image through the logarithmic transformation to the logarithm domain,so that you can low-dose perspective image of the multiplicative noise into additive noise for processing.The second part extends the construction of the training sample set to a similar image set(generally select two images from top to bottom).In this process,it is also proved that the image containing the noise is the same as the covariance matrix without noise images Of the eigenvector.The experimental results show that the algorithm can effectively protect the details of the image while reducing the noise.In addition,this paper also improves the noise reduction algorithm based on nonlocal mean:non-local mean algorithm(SINLM)based on similar image.Similar to the PCA-based improved method,the search for similar blocks is extended to other similar images.The improved algorithm greatly improves the possibility of finding similar blocks and overcomes the shortcomings of non-local mean algorithm over-reliance on the structure information of the image itself.In the noise reduction while reducing the smoothness of the image.As an important parameter of the classical non-local mean algorithm,the algorithm of improving the setting of the weight parameter h is also discussed again.Since the other images are introduced to assist in the processing of the current image,the improved algorithm will be similar to the current image The image is introduced into the parameter h,and the formula of h is given.