Neural network regularization in the problem of few-view computed tomography

The computed tomography allows to reconstruct the inner morphological structure of an object without physical destructing. accuracy digital image reconstruction directly depends on measurement conditions tomographic projections, in particular, number recorded projections. In medicine, reduce dose patient load there try measured However, a few-view tomography, when we have small using standard algorithms leads reconstructed images degradation. main feature our approach for is that algebraic being finalized by neural network with keeping projection data because additive result zero space forward operator. final presents sum calculated and reconstruction. First element second orthogonal addition space. Last applying method few-angle sinogram. dependency model between elements operator built networks. It demonstrated realization suggested achieving better computation time than state-of-the-art approaches test from Low Dose CT Challenge dataset increasing reprojection error.