Pre-reconstruction restoration of SPECT projection images by a neural network

In single photon emission computed tomography (SPECT) the projection images obtained at view angles surrounding the patient are degraded due to the geometric response of the imaging system (a spatially-variant blur), Compton scatter, Poisson noise, and other factors. Various methods have been proposed for compensating for the spatially varying geometric response of the camera. In this study the authors examine restoration of SPECT projection images using an artificial neural network. A three layer feed-forward neural network is trained to compute the spatially-variant standard deviations of a symmetric Gaussian blur. A Hopfield network is then used to restore the projection images in which the restoration problem is formulated as a minimization of an error function of the network. Results from applying this restoration procedure on SPECT projection images are presented and the resulting SPECT reconstruction are analyzed