A neural network approach to real-time discrete tomography

Tomography deals with the reconstruction of the density distribution inside an unknown object from its projections in several directions. In Discrete tomography one focuses on the reconstruction of objects having a small, discrete set of density values. Using this prior knowledge in the reconstruction algorithm may vastly reduce the number of projections that is required to obtain high quality reconstructions. Recently the first generation of real-time tomographic scanners has appeared, capable of acquiring several images per second. Discrete tomography is well suited for real-time operation, as only few projections are required, reducing scanning time. However, for efficient real-time operation an extremely fast reconstruction algorithm is also required. In this paper we present a new reconstruction method, which is based on a feed-forward neural network. The network can compute reconstructions extremely fast, making it suitable for real-time tomography. Our experimental results demonstrate that the approach achieves good reconstruction quality.