A Sparse and Ergonomic Tomographic Image Reconstruction Technique based on Artificial Neural Networks

Considering the symmetry properties exhibited in the Radon Transform, a new type of tomographic image reconstruction technique based on conventional Feed-Forward type Artificial Neural Networks with fully equipped multilayer perceptrons is presented in this work. By introducing a simple remapping of the original sinogram in a constructive and sparse way without loosing the basic projective information, an ergonomically improved network architecture is here presented. The proposed prototype is free of any training sample dependencies or correlations. It provides comparable results to the standard reconstruction algorithms by reducing the computational time of the process. A variety of successfully reconstructed images with software phantoms but also with clinical SPECT data are shown and discussed.