Evaluation of a neural-network classifier for PET scans of normal and Alzheimer's disease subjects

Abstract

The value of PET as an objective diagnostic tool for dementia may depend on the degree to which abnormal metabolic patterns can be detected by quantitative classification methods. In these studies, a neural-network classifier based on coarse region of interest analyses was used to classify normal and abnormal FDG-PET scans. The performance of neural networks and of an expert reader were evaluated by cross validation testing. When the "abnormal" class was represented by subjects with clinical diagnoses of "Probable Alzheimer's," the areas under the relative-operating-characteristic (ROC) curves were 0.85 and 0.89 for the neural network and the expert reader, respectively. When testing with abnormal subjects represented by "Possible AD" cases, ROC areas for both the network and the expert were 0.81. The neural network out-performed discriminant analysis. It is concluded that PET has potential for the detection of abnormal brain function in dementing diseases, and that the combination of neural networks and PET is a useful diagnostic tool. Despite the low-resolution "view" afforded the neural network, its performance was nearly equivalent to that of an expert reader.