An artifical neural network using input data derived from attenuation measurements was trained to generate spectral profiles (relative number of photons versus energy). Once the relative spectral distribution is reconstructed, absolute spectra (number of photons per unit exposure versus energy) can be calculated. A neural network was trained on spectra generated mathematically using the Birch-Marshall model, combined with attenuation data, calculated from the spectra by munerical integration. Wheras attenuation data can be calculated in a straightforward manner from the x-ray spectra, the reverse is not true. Several neural networks were successfully taught to reconstruct the spectra, given the attenuation data. The networks were tested using kV/inherent filtration combinations that were not in the training set, and the performance of the reconstruction was excellent. Noise in the attenuation data was simulated to test the effects of noise propagation in the reconstruction. The effects of network architecture and data averaging on noise propagation were investigated. Experimentally determined spectral data compiled by Fewell were also used to train a neural network, and the results of the reconstruction were also found to be excellent.
ASJC Scopus subject areas