Neural networks for parameter identification in compartment models

In health physics and nuclear medicine, numerous situations arise in which it is required to model the transport of radionuclides in biological systems. Compartment models and Patlak graphical analysis methods have been employed for quantification of these parameters. This thesis evaluates the application of artificial neural network models to identify model parameters in such applications.

Neural network models were developed for predicting various model parameters. A trial and error procedure was employed to obtain a suitable number of hidden nodes and hidden layers to achieve appropriate network structure for health physics and nuclear medicine applications. Computer algorithms were written to generate training and testing data for these network models. After the networks were trained, their performance was evaluated on unseen patterns (not used in training) of synthetic data by comparison of results from the artificial neural network models with those obtained from traditional methods.

The neural network models trained with synthetic data were also tested with experimental data obtained from Positron Emission Tomography (PET) scans. In addition, several neural network models were developed and trained using different mathematical features of the synthetic data. Then, the performance of these network models were evaluated by testing with unseen patterns (not used in training) of the synthetic data.

The results show that the network models predicted the model parameters with more than 95% accuracy. Myocardial blood flow (MBF) estimated by neural network models correlated well with MBF obtained by Patlak method (correlation (r) 0.97, slope 1.0055). MBF obtained by neural network models agreed well with MBF obtained by those specified by the compartment models (r = 0.98, slope = 1.01). Good agreement of the MBF estimates from neural network models were also observed for data containing small amounts of noise (r = 0.97, slope 1.0092).

Results from these studies demonstrate that neural network models are reliable and suitable tools for identification and estimation of various parameters in compartmental models. In addition, the neural network models can capture the nonlinear characteristics of data more efficiently and in less complicated manner than traditional techniques.