Sparse-view CT Reconstruction Algorithm Based On Transformer

Sparse-view computed tomography(CT)is an effective approach for reducing radiation dose and speeding up scanning,but it often leads to significant streak artifacts.Traditional model-based iterative reconstruction(MBIR)methods have good anti-streak performance,but sensitive hyperparameter selection and manually designed priors are major obstacles in clinical applications.In recent years,deep learning(DL)methods have been widely used due to their excellent computational efficiency and reconstruction performance.However,traditional CNN models cannot fully capture global contextual dependencies in structural representations in sparse sampling scenarios due to the inherent locality of their convolutional kernels,leading to information loss and decreased reconstruction quality.Additionally,DL-based sparse-view CT reconstruction methods generally suffer from insufficient generalization ability,limiting their clinical potential.To address the existing problems in DL-based sparse-view CT reconstruction methods,this paper introduces Transformer model and score-based generative model,aiming to leverage the global information capture capability of Transformer model and the powerful generation capability of fraction-based generative model to further improve reconstruction performance and generalization ability.The primary contributions of this research include:(1)To address the inefficiency of existing convolutional neural network-based methods in modeling CT images with various structural information,this paper proposes a dual-domain sparse-view CT reconstruction algorithm based on Transformer.This algorithm integrates dual-domain reconstruction into a unified network and employs Swin Transformer as the primary building block to enhance the extraction capability of global projection and image features.Qualitative and quantitative experimental results demonstrate that this algorithm has significant advantages in enhancing reconstructed image details and improving image accuracy,validating the potential of Transformer in sparse-view CT reconstruction.(2)To address the over-smoothing problem in ultra-sparse-view reconstruction using existing deep learning methods,this paper proposes a sparse-view CT reconstruction algorithm based on Transformer frequency-domain supervision.This algorithm introduces a frequency-domain supervision module on top of dual-domain reconstruction to fill in missing coefficients in the Fourier domain,successfully reducing the number of projection angles required for reconstruction.Additionally,it incorporates a sparse attention mechanism and a mixed-scale feedforward network to improve the standard Transformer architecture,aiming to reduce computational load while enhancing reconstruction performance.Qualitative and quantitative experimental results demonstrate that this algorithm exhibits superior performance in ultra-sparse-view CT image reconstruction,effectively balancing the reduction of projection angles and image reconstruction quality.(3)To address the reliance on large paired datasets for training and the insufficient generalization capability of existing deep learning methods,this paper proposes a sparse-view CT reconstruction algorithm based on a Transformer score-based generative model.This algorithm utilizes a score network based on the Swin Transformer architecture to learn the global prior knowledge of images,more accurately modeling the noise and artifacts in the global distribution of sparse-view CT images.By integrating the optimization strategy of MBIR into the diffusion sampling steps,this model achieves a seamless fusion of data-driven deep generative priors and traditional priors,thereby enhancing the algorithm’s generalization capability.The completely unsupervised reconstruction algorithm avoids dependence on large-scale paired data,alleviating the issue of relatively limited training samples in medical datasets.Qualitative and quantitative experimental results demonstrate that this algorithm not only achieves reconstruction results comparable to supervised learning methods but also exhibits good generalization capability.