Research On Low-Dose Problems Of Micro-CT Via Dictionary Learning

Micro Computed Tomography (Micro CT) is widely been used in clinical small-animal model tests and micro-structure research due to its high resolution, non-invasion and fast imaging. According to the imaging principles, there has always been existing contradictions between the high resolution of the system, the focus size of the micro x-ray source and the integration time of the detector. To guarantee the quality of the reconstructed image, Micro CT needs a long scan to accumulate enough dose of x-ray which has a bad effect on in-vivo biological samples. The decrease of the x-ray dose will cause a significant degradation on the reconstructed image. Therefore, the problem of Micro CT with low dose attracts wide research attentions.There are mainly two methods applying on x-ray dose reduction currently:either the dose reduction on single projection or number of total projections decreasing. The two methods both can be regarded as the absence of the reconstructed information, which will cause the noise model change, the Signal to Noise Ratio (SNR) degradation and the artifacts introduction (e.g. ring artifacts, hardening artifacts and low-dose stripe artifacts). Based on the compressed sensing theory, the image signal can be reconstructed with less information while utilizing the sparsity of the signal itself. This kind of methods will be very suitable to solve the information deficient problems caused by low-dose CT. As a branch of the compressed sensing theory, the dictionary learning is very sensitive to the structural information, can make the image signal a good sparse decomposition. On this basis, this paper will focus on three aspects, which are respectively the SNR improvement of the reconstructed images, the iterative reconstruction of sparse projections and the reduction of the artifacts, to explore the image quality issues caused by low-dose of Micro CT.In the course of the SNR improvement, the dictionary learning is used as a denoising method to improve the image quality. In the ordinary sense, the dictionary learning from the noise image can get a better performance on denoising, but the procedure for each single reconstructed image will cost a vast of time. Accordingly, this paper proposed a global dictionary which performs learning form a high-dose reconstructed image. With the global dictionary, the denoising process can save the time of dictionary learning efficiently while maintaining the denoising performance. The real data experiments demonstrated that, compared with the mainstream denoising algorithms, the proposed method can retain more image details while denoising.In the research of sparse projections reconstruction, based on assumption of the priors that the image signal can be sparsely decomposed, the dictionary is introduced into the iterative process as a penalty term. The proposed method can sparsely decompose the image signal during iteration and reduce the number of equations. In addition, with the intrinsic denoising ability of the dictionary, the iteration based on dictionary learning can get a better performance on sparse projections reconstruction. The sparse projections reconstruction experiments validated the effectiveness of the proposed method.For the different kinds of artifacts introduced by low-dose reconstruction, this paper focuses on the ring artifacts and the low-dose stripe artifacts. The ring artifacts share some similar structures with the scanned objects on cross-section, filtering only on cross-section often resulted in the loss of image structural information. This paper took discontinuity of the ring artifacts in spatial space into consideration, proposed a filtering method based on three dimensional dictionary. On the other hand, in view of relatively single characteristics of the structural features that the low-dose stripe artifacts demonstrate, this paper built a dual-dictionary method which contained a structural dictionary and an artifacts dictionary. With the dual-dictionary, the low-dose stripe artifacts can be filtered out during the image signal reconstruction stage. Experiments showed the dictionary based artifacts filtering algorithms could filter out the artifacts effectively without introducing blurry and new artifacts.After all, considering that all algorithms mentioned above had a high computational cost, this paper applied those ones on GPU platform to achieve the feasibility for the applications.