Unsupervised learning of spect reconstruction

An approach of image reconstruction from projection in single-photon emission computed tomography (SPECT), based on an unsupervised learning artificial Kohonen neural network, is developed. A kind of random sampling technique is used to generate sensory input from tomographic projections. It is proved that an adequate interpretation and representation of the synaptic strengths of the network can be used to obtain the tomographic image. A relevant numerical experiment is reported. To validate the method tests were performed using a typical phantom (simulated object) used in limited data sets tomography. At the end of the self-organizing process the synaptic strengths can be shown on x-y plane as a map of points. The density of points gives the reconstructed image. In order to obtain a usual representation of the image a grid of 32 x 32 pixels was superimposed on the map and the number of points in each pixel was counted in order to form the image function. As the number of points in each pixel (especially for pixels corresponding to the background of the phantom) is relatively small, the image function is spoiled by noise. Noise filtering was used to improve the quality of the image. The final result is presented as a 3D-surface plot and as a corresponding 8-level gray map. It can be concluded that a good reconstruction of sizes and shapes was obtained. Some artifacts spoil the region located around the spikes. This artifact can be reduced if an increased number of external stimuli are presented to the network. The price paid for this improvement is the increase of computation time. (authors)