Research On Image Reconstruction And Artifact Correction Methods For Low Dose CT

CT is one of the most commonly used medical equipment in the process of disease diagnosing and treatment. In many diseases CT can not be replaced by other imaging devices. However, as a kind of radiation equipment, widespread usage of CT increases X-ray absorption dose of the patients, which may cause potential harm to their health. Since low-dose CT can reduce X-ray to the patients, it attracts more and more attention and study, aiming at achieves good quality of reconstructed images. Reducing the radiation dose can be achieved by methods like reducing the scanning angle, reducing the scan time and using lower radiation current. Among these methods, reducing the scanning angle leads to incomplete projection data, while iterative reconstruction algorithms can effectively solve the problem. Using lower radiation current or short scan time will cause larger noise ratio in the projection data due to less number of photons, results in Gaussian noise artifacts in the reconstructed image. It is an effective way to solve the problem of image and projection noises to perform denoising in image and projection domain. For patients with metal implants in CT scanning, metal artifacts will be generated in reconstructed images. How to recover the affected metal projection data accurately is a very significant issue. Abnormal detector units produce ring artifacts in the reconstructed image. Filtering in projection domain can effectively correct ring artifacts. This dissertation mainly focuses on the key factors and the difficult points of low dose CT image reconstruction and artifact correction. Research works were carried out as follows.(1) In order to reconstruct high quality CT images from imcomplete projection data, this dissertation presents an ordered subsets reconstruction algorithm. The algorithm uses projections onto convex sets to accelerate the rate of reconstruction convergence, combines total variation minimization and fast first order method to reduce the iteration times of reconstruction, and uses split Bregman alternating direction method to solve optimization problems. Simulation data and real data experiments show the reconstruction results of the proposed methods when using incomplete projection data. The proposed algorithm provides both a theoretical and experimental basis for the practical application.(2) To reduce Gaussian noise of CT images, the author proposed a CUDA-based accelerated three-dimensional total variation minimization algorithm. The algorithm automatically calculates the iterative step size and reduces the total variation of the images by using gradient descent method. Experimental results of simulation data and real data were analyzed, in order to verify that the proposed accelerating algorithm can effectively reduce the computation time and most Gaussian noise. The algorithm can maintain the texture and boundary information of the images.(3) For the problem about noise of low dose CT projection data and parameter selection of denoising model, according to the different denoising stages in projection domain, this dissertation presents four denoising algorithms based on before and after logarithmic transformation. In the projection data denoising process before the logarithmic transformation, this dissertation proposes a gradient descent method with adaptive step size based on dimension increase of one-dimensional projection data or two-dimensional projection data. This dissertation also presents a method to automatically calculate the penalty parameters of the alternating direction method of multipliers before the logarithmic transformation. In the projection data denoising process after the logarithmic transformation, according to the noise model of projection data after the logarithmic transformation, this dissertation presents a denoising algorithm that can automatically calculate noise variance. The algorithm uses the variance stability transform to convert signal-dependent Gaussian noise into signal-independent Gaussian noise, selects a noise variance evaluation method to determine the denoising parameters, uses block-matching and 3D filtering, and referenced non-local means to reduce noise of the images. Experimental results of the simulated and real data show the feasibility and effectiveness of the proposed four algorithms.(4) For reducing metal artifacts of CT images, this dissertation proposes a correction algorithm based on redundant representation. First, the areas of metals were segmented from CT images and projected to obtain a metal projection data. Then, the projection data of the metal parts were removed from the original projection data. Finally, the projection data of metal parts were recovered using redundant representation method. Corrected images were reconstructed using filtered back projection method. Corrected results show that the proposed algorithm can effectively reduce metal artifacts of the reconstructed images.(5) For reducing ring artifacts of CT images, this dissertation proposes a correction algorithm based on CUDA. The method uses GPU accelerating median filtering process of two-dimensional projection data. Corrected results show that the proposed method using GPU is faster than the CPU processing method and eliminates the ring artifacts while maintaining the spatial resolution of the images.