Chapter 54 - Parameter Estimation of the Fluorodeoxyglucose Model with a Neural Network

Abstract

An artificial neural network (ANN) is a trainable algorithm that can produce an output appropriate for a given input. Such networks can be applied in a wide variety of pattern recognition tasks, including parameter estimation. The potential advantages of using ANNs for parameter estimation are speed and noise tolerance. The parameter estimation task studied is the five-parameter fluorodeoxyglucose (FDG) model (K1, k2, k3, k4, and Vp). Training and test data sets are generated from a program that produces both blood time–activity curves (TACs) and tissue TACs using the FDG model. The parameters of both the blood and FDG models are randomized to produce 1,000 data sets for training and six separate 1,000 data sets for testing. One test data set is noise free and the others had 5, 10, 15, 20, and 30% SD normally distributed noise injected into the tissue TAC only. The ANN was a three-layer network trained with back propagation. The input consisted of 20 points from the blood TAC and 20 points from the tissue TAC. Differing numbers of hidden nodes were tested. Best results are obtained with three hidden nodes, estimating only glucose metabolic rate. Performance was considerably better than with graphical analysis. This suggests that ANNs may be an effective approach to generate pixel-by-pixel functional images.