Accurate scatter compensation using neural networks in radionuclide imaging

Abstract

A method to estimate primary photons using an artifical neural network in radionuclide imaging is presented. The neural network for /sup 99m/Tc has three layers, i.e., one input layer with five units, one hidden layer with five units, and one output layer with two units. As input values to the input units, count ratios are used. They are the ratios of the counts acquired by narrow windows to the total count acquired by a broad window with the energy range from 125 to 154 keV. The outputs are a scatter count ratio and a primary count ratio. Using the primary count ratio and the total count, the primary count of the pixel is calculated directly. The neural network is trained with a backpropagation algorithm using calculated true energy spectra obtained by a Monte Carlo method. The simulation shows that an accurate estimation of primary photons is accomplished within an error ratio of 5% for primary photons. >