Аннотация:In this work, we proposed a new complex algorithm, consisting of solutions to three problems that were solved using neural network models.The first problem was the projection noise, which is caused, in particular, by X-ray probe instability. There was used our residual convolutional network for noise reduction on projections. It is shown that the proposed approach is computationally efficient in terms of the ratio of quality to running time compared to previously proposed algorithms.The second issue under consideration is to increase the speed of reconstruction. A time-efficient image reconstruction algorithm from low noise projections was proposed. It is a modification of the FBP algorithm, in which the frequency filter is selected by the machine learning method in accordance with the dependency to the projection measurement geometry. This allows us to get the same reprojection error as in computationally complex iterative algorithms, such as, for example, SIRT.The third alleged problem is the regularization of solutions with zero reprojection error in the case of a small number of projections. In this case, there is more than one image, which is conditioned to the measured projections. To select the correct image from all possible, we proposed the residual convolutional network to SIRT reconstruction with interpolation and normalization in the Fourier surface.