Optimization and evaluation of a neural network classifier for PET scans of memory disorder subjects

Abstract

Back-propagation neural networks were used to classify PET scans as either normal or abnormal, with abnormal subjects defined as subjects who had previously been clinically diagnosed with memory disorders. Numerous neural network experiments were performed in order to achieve optimization with respect to number of hidden units and training duration. Optimizations and performance evaluations were based on ROC analysis, in which the area under the ROC curve was the figure of merit. The neural network's performance was better than that of dlscrlminant analysis, and comparable to the expert's performance, despite the low resolution image data, which consisted of one value per brain lobe, provided to the network.