System Modeling Based Low Dose CT Reconstruction

X-ray computed tomography (CT) has become an excellent example for medical imaging with its high temporal resolution, spatial resolution and contrast resolution. It has been widely used in the clinical diagnosis and treatment. But the excessive X-ray radiation may induce cancer, leukemia, or some other hereditary disease, so the safety and security of CT imaging has become one focus problem in the industry. How to get the best CT image quality with the minimum cost and minimum X-ray dose, has become one of the most important issue.Until now, various techniques have been investigated, including directly lowering the mAs or down-sample the viewing angle, to reduce radiation dose in CT scans. Meanwhile, the associated reconstructed image usually suffers from serious noise and artifacts, which has negative influence for clinical diagnosis. Recently, investigation focused on low dose CT reconstruction can be grouped into three categories:the post-processing of the reconstructed CT image, low dose CT projection restoration, and iterative reconstruction methods that include the algebraic iterative reconstruction algorithm and statistical iterative reconstruction algorithm. Among them, the study on iterative reconstruction algorithms is more popular. Compared with the analytical reconstruction algorithm, the iterative reconstruction algorithm can overcome the inherent physical limitations of CT by modeling the imaging geometry, X-ray spectrum, beam hardening, scattering, and noise, etc. Therefore, the iterative reconstruction algorithm can further improve the image spatial resolution and reduce the artifacts. Furthermore, iterative reconstruction algorithms that based on photon statistics constructing an accurate noise model, has a better performance in reduce image noise. Generally, accurate modeling of projection data measurements is the foundation of high quality reconstruction, and the introduction of prior knowledge to avoid solving ambiguity and high precision image reconstruction has very important significance.X-ray CT technologies have been widely explored for specific applications in clinic including perfusion imaging,4D-CT imaging, image-guided intervention and radiotherapy, et al. Under these situations, repeated topographic acquisitions are often prescribed. For instance, except the planning CT, in daily Cone Beam CT (CBCT) examinations for target localization in image-guided radiation therapy (IGRT), repeated scans have become routine procedures. In this case, the cumulative radiation dose still significantly increase as comparison with the conventional CT scans, which has raised major concerns in patients. With regard to the repeated CT scans, a previously scanned high-quality diagnostic CT image volume usually contains same anatomical information as the current scan except for some anatomical changes due to internal motion or patient weight change. In other words, there exists redundant information among the repeated scan CT images. Traditionally, the prior information is constrained on the different values of the local neighboring pixels in the image domain. So the CT scans at different times are often dealt independently and no systematic attempt has been made to integrate the valuable patient-specific prior knowledge, i.e., the previous scanned data that hold a wealth of prior information on the patient-specific anatomy, to promote the subsequent imaging process. According to the above, in this paper, we focus on exploiting reasonable approaches to incorporate the redundancy of information in the prior-image and also the optimization with the introduction of the prior information. Meanwhile, we have a further study on the noise prosperities of the flat panel detector (FPD) mounted on the CBCT imaging system, which has constructed the statitiscal model for CT iterative reconstruction. On the whole, the main works of this paper can be summarized as follows:1) To introduce the prior information in a reasonable way, and to overcome the disadvantage of the locally designed prior term with only the different values of the neighboring pixels as the constraint, in this paper, we propose to utilize the nonlocal criteria to search the patched based similarity between different pixels to form the regulation term. In this way, the prior information can be introduced efficiently. The new approach relaxes the need of accurate registration between the target image and the prior image. Experiments on the physical phantom, numerical XCAT phantom, and patient data show that the proposed approach can improve the image quality without introducing new fake structure in low dose CT imaging.2) Base on two main characteristics of cerebral perfusion CT imaging:(1) the normal dose pre-contrast scanned image has relative higher resolution and lower noise level as compared to the follow up low dose contrast image. So the redundant information in the precious high quality image can be used to promote the low dose contrast image reconstruction.(2) During the sequential scan of the selected slices, most of the anatomical structures are unchanged except the changes of the intensity level. In his paper, we propose a normal dose pre-contrast scan induced low dose cerebral perfusion CT image reconstruction algorithm. Firstly, the low dose contrasted image is restored by using the improved nonlocal means criterion with the normal dose pre-contrast image. Then the image is reconstructed by an iterative deconvolution reconstruction framework. Experiments on both numerical simulation and patient data show that the algorithm can effectively suppress noise and improve the image SNR, and then to make the corresponding hemodynamic parameters calculation more accurate.3) To address the mismatch problem between the prior image and target image for the prior image constrained compressed sensing (PICCS) algorithm, in this paper, we propose to obtain similar or closed image volume by registering the on-treatment projection data and the prior image volume. Particularly, this procedure is facilitated by estimating the deformation vector fields (DVF) through matching the forward projection of the prior image and the measured on-treatment projection. Then by translating the DVF onto the prior image, we get the deformed-prior image that used as the prior image for prior-image based image reconstruction algorithm. The present approach can effectively avoid the negative influence of the noise and artifacts in the FDK reconstructions in the image domain registration. Experimental studies on the XCAT phantom and patient data show that the proposed approach generates high-quality registered prior image with most anatomical structures aligned to the target image, the image quality of the reconstructions has been improved as compared to the standard PICCS algorithm.4) Given the physically difference between the flat panel detector in CBCT imaging system and the conventional fan beam CT detector, in this study, we systematically investigated the noise correlation properties among detector bins of CBCT projection data on a True-Beam on-board system. The noise correlation coefficients among detector bins were calculated with repeated scanned measurements. The analyses showed that non-zero noise correlation coefficient values can be found between the eight nearest neighboring pixels. For the first order neighbors, the noise correlation coefficients are larger than the second order ones. Meanwhile, the noise correlation coefficients are independent of the dose level. The noise correlation among CBCT projection data was then incorporated into the covariance matrix of the penalized weighted least-squares (PWLS) criterion for noise reduction. Reconstruction shows that the consideration of noise correlation results in improved reconstruction quality. An accurate noise model of CBCT projection data was obtained.