基于双域深度神经网络的LACT重建方法研究OA

Limited-Angle Computed Tomography(LACT)is designed to reconstruct the original CT images using angularly restricted projection data.Images reconstructed by conventional methods contain severe artifacts or even distortions due to the incomplete projection data.Deep learning-based methods can address this deficiency,however,the deep neural networks constructed by existing deep learning-based LACT methods are usually empirically designed,so the model architecture is not interpretable.In addition,the existing methods do not fully utilize the projection domain information for network training,which leads to the reconstruction accuracy to be improved.In order to solve these problems,a dual domain reconstruction optimization model for joint image and projection domains is constructed from the dual domain perspective of CT images and projection data,the model is solved by using the proximal gradient descent algorithm,and the iterative steps are unfolded into a deep neural network,which constructs a dual domain deep unfolding network for limited-angle CT reconstruction.The simulation results show that the PSNR of this dual-domain deep unfolding network reaches 27.70 dB,30.17 dB and 33.98 dB at limited angles of 90°,120° and 150°,respectively,which outperforms the existing mainstream deep-learning-based methods.In addition,the CT images reconstructed by the deep unfolding network not only remove artifacts but also preserve more of the image's tissue structure and detail information,achieving excellent visual results.