Learned Full-sampling Reconstruction From Incomplete Data

Computed tomography(CT)projects the sample from different angles by X-ray,and then uses mathematical models such as wave back projection,algebraic method or statistical method to reconstruct the image.Sparse-view CT reconstruction and Limited-angle CT reconstruction are both very challenging problems in practice.Due to the high morbidity of incomplete data reconstruction,both the analytical reconstruc-tion method and the iterative reconstruction method may lead to the distortion and arti-facts of the reconstructed image.This paper mainly introduces a new CT reconstruction model,which can estimate a high resolution Radon domain data while reconstructing a high quality reconstructed image.This model is obtained by regularization of CT images and projection data at the same time,and according to the full-sampling con-dition,the full-sampling system matrix is introduced to fit the reconstructed image and the high-resolution projection data.Inspired by the success of the deep learning method in the imaging field,and combined with the ADMM algorithm of the model,we de-signed an end-to-end convolutional neural network to learn the relationship between the original data and the reconstructed image,as well as the prior information in the data.The high quality reconstructed image and high resolution data can be obtained by inputting the projective domain data into the trained convolutional neural network.The experimental results show that our model is more suitable for sparse-view and limited-angle CT reconstruction than variation-based CT reconstruction and learning-based CT reconstruction.