| Image reconstruction is a technology to obtain three-dimensional object shape information by digital processing from the external measurement data.The essence of CT technology is to emit x light from an x-ray source,hit the detector through a three-dimensional object,and perform a series of processing by acquiring the data received on the detector to obtain external measurement data corresponding to the object.CT image reconstruction is to reconstruct the two-dimensional slice of the three-dimensional image from these processed external measurement data and to generate a tomographic image.Doctors can deal with the clinical diagnosis accurately,efficiently and reliably by observing the reference CT images.At the same time,CT image noise pollution can not be avoidable due to the imaging principle and equipment impact.CT image noise in the uniform material image is expressed as the difference between the CT values and its mean values in a given region.Doctors can not diagnosis accurately and rapidly because the image resolution is reduced and some diseased tissues in the image are blurred by noise.This paper focuses on the sparse view CT medical image reconstruction based on MSTV and the de-noising algorithm based on high order singular value decomposition(HOSVD).The important role of regularization in CT medical image reconstruction is discussed emphatically.We introduce a new regularization,namely,Mumford-Shah total variance(MSTV),and introduce MSTV into the framework of penalty weighted least squares.This paper also obtains a new reconstruction model by integrating MSTV into the tranditional energy function of CT medical image reconstruction.In order to reduce the influence of sparse view data acquisition,the high order singular value decomposition technology is introduced into the CT image denoising algorithm.The high order singular value decomposition and transform operation of the CT image in the reconstructed domain are employed in this paper.The gray value of the pixel after clustering is obtained by the weighted average method in the transformation domain.A quantitative and qualitative result analysis about the resolution and accuracy is made on computer simulation digital phantom in this paper.The contributions of this paper are as follows:1.A sparse view CT image reconstruction method based on MSTV is proposed.TheMSTV regular term and the penalty weighted least squares constraint are integrated into the same framework,and a new energy function model is constructed.In view of the dual regularity of MSTV,the new model preserves the edge information well while reconstructing the image.The new model includes the parts of image reconstruction,image denoising and edge extraction.These three parts are not isolated from each other,the results of the previous step affect the results of the latter step directly.In addition,the three-step optimization can also be more flexible to deal with CT images in different situations,for example,the denoising process in the third step can be repeated several times when CT image with heavy noises.Compared with the classical sparse view CT image reconstruction model based on the total variation minimum,experimental results show that the proposed model can obtain better reconstruction results and sharpen the edge contours.2.A CT image denoising algorithm based on HOSVD is proposed.The challenge of image denoising is to remove the noise while preserving the edge information.The non-local image denoising method,which is based on the sparseness and structural similarity between the image blocks,can achieve good denoise results while maintaining the image detail information recently.However,the basis is fixed and not flexible in traditional methods.In this paper,we propose a new CT image denoising algorithm based on HOSVD.The new algorithm can select the appropriate basis by the image adaptability learning.The new algorithm separates the reconstructed images into several groups firstly,and searches blocks matching the reference blocks in the surrounding window for any reference block in each group,and stacks the similar block groups in a three-dimensional matrix,and then the high-order singular value decomposition of the two-dimensional block in the three-dimensional matrix is transformed to obtain the coefficient and basis.The hard threshold operation is employed to select the transformation coefficient and the inverse transformation of high order singular value decomposition is also applied finally in the algorithm.The above operations are repeated for each of the reference blocks,and the final denoised image is obtained by weighting average the estimated values after multiple processing. |
