Generative Model In Sinogram Domain For Sparse-view CT Reconstruction

X-ray radiation in Computed Tomography(CT)examinations can cause harm to patients,but the total radiation dose during the examination can be significantly reduced by reducing the projection angle.However,directly reducing the projection views often leads to severe aliasing artifacts in the reconstructed image.To address the imaging artifacts caused by reduced projection views,various reconstruction methods have been proposed,including those based on deep learning.In the past,deep learningbased sparse-angle CT reconstruction required supervised training of the network using sparse-view/full-view CT image pairs.When the number of projection views changes,supervised deep learning-based reconstruction methods require retraining the network using corresponding new sparse-view/full-view CT image pairs to maintain the reconstruction quality.This makes these methods inflexible in handling various sparseview CT reconstruction tasks.To alleviate this limitation,this paper proposes a completely unsupervised generative model based on sinogram domain score matching for sparse-view CT reconstruction tasks.By applying the channel duplication strategy,irrelevant noise is injected into the samples during training and reconstruction,which improves the accuracy of score estimation and reconstruction performance.Finally,this paper qualitatively and quantitatively evaluates the proposed method on multiple relevant CT datasets,and the experimental results show that this method achieves comparable or better performance than supervised methods.The main contributions of this study are as follows.(1)A score-based generative model based on the high-dimensional tensor of the sinogram domain is designed.By using the channel replication strategy,the original projected data is mapped into a high-dimensional tensor to improve the accuracy of score estimation based on the generative model based on score matching.Then,multiscale noise is used to disturb the constructed high-dimensional data to estimate the gradient of its probability distribution density.Then,in the projection generation stage,the inverse stochastic differential equation solver and data consistency constraint are used to update the fidelity and prior terms alternately to ensure the optimal solution.(2)A channel-copy strategy is proposed to construct a high-dimensional sinogram domain tensor,which improves the accuracy of score matching and achieves a good performance in sparse views CT reconstruction applications.At the same time,it also alleviates the problem of insufficient representation ability of the generated model in processing complex data distribution data.