Efficient Neural Network Image Reconstruction from Raw Data Using a Radon Inversion Layer

Direct neural network image reconstructions methods from raw data are notoriously memory inefficient in the transformation from the measurement space domain to image space by use of one or more fully connected layers. This inefficiency has thus far limited the size of of the final image and posed a significant obstacle to studying direct neural network reconstruction methods. However, this memory space challenge can be corrected by observing the measurement data that contributes to the formation of a given image patch and designing a network layer to exploit those observations. This article takes this approach and proposes a novel Radon inversion neural network layer that enables full size direct image reconstruction from Radon encoded raw measurement data with a 99.84% decrease in memory requirements compared to the recently published AUTOMAP direct reconstruction network. We demonstrate high quality full size image reconstruction from a neural network using the Radon inversion layer and compare the results to the SART reconstruction method utilizing a structural similarity measure for comparison.