An artificial neural network for tomographic reconstruction

摘要: Tomography became increasingly popular in the last years in scientific fields like optics, fluid flow, plasma diagnosis, non-destructive testing. The main characteristic that set the problem apart from that encountered in nuclear medicine is the limited data sets (the small number of projections, the confinement of the projection angles in a narrow range). To deal with these constraints we suggest the use of a three-layer feed-forward artificial neural network. The number of synaptic connections is minimized by taking into account the geometry of the tomographic problem. The connections between the input layer and the hidden one perform the filtering of the experimental data and the connections between the hidden layer and the output one (which is the reconstructed image) model the back-projection process. The performances of the method are evaluated with a numerically generated phantom in a geometry characterized by eight projection angles, limited in the range 0 angle to 76 angle. The quality of the reconstruction is evaluated by comparing to a conventional method that uses the statistical criterion of maximum entropy. (authors)