Artificial neural networks for SPECT image reconstruction with optimized weighted backprojection

An artificial neural network has been developed to reconstruct quantitative single photon emission computed tomographic (SPECT) images. The network is trained using known projection-image pairs containing Poisson noise to learn a shift-invariant weighting (filter) which minimizes the mean squared error between the reconstructed image and the sample image. Once trained, the network produces a reconstructed image as its output when projection data are presented to its input. Supervised training with a modified delta rule was used to train this two-layer neural network having a backpropagation architecture with one hidden layer (the filtered projection). The system was trained for noise levels representative of 1000, 10k, and 1M counts per slice. The Fourier transform of the filtered projection (for an impulse function) is compared with the ramp filter used with backprojection.<>