Simulation Study on Kinetic Parameter Estimation Using Artificial Neural Networks.

概要

論文の詳細を見る
<I>We applied artificial neural networks (ANNs) to kinetic parameter estimation and investigated their performance using a computer-based simulation method. The ANNs used in the present study were based on a feed forward layered model with an input, hidden, and output layers, on which a backpropagation learning model was implemented. In the computer simulation, the three-compartment fluorodeoxyglucose model was used. The brain time-activity data were generated using the mean and standard deviation of the rate constants in the literature and the plasma time-activity data obtained from actual positron emission tomography studies. These time-activity data and the rate constants were used as input and output data for training ANNs, respectively. The pe rformance of ANNs was generally better with increasing number of hidden nodes and training samples. Although the accuracy of the parameter estimates obtained by ANNs was lower or comparable to that obtained by the nonlinear least-squares (NLSQ) method at lower noise levels, it was generally better than that obtained by the NLSQ method at higher noise levels. Although the computational cost for training ANNs was very high, ANNs could estimate the parameters very fast once they were trained adequately. These results suggested that an ANN approach may be an effective strategy for kinetic parameter estimation</I>.
社団法人日本アイソトープ協会の論文