Neural network classification of normal and Alzheimer's disease subjects using high-resolution and low-resolution PET cameras

Abstract

Neural-network classification methods were applied to studies of FDG-PET images of the brain acquired from a total of 77 "probable" Alzheimer's disease and 124 normal subjects at two different centers. Methods: Classification performances, as determined by relative-operating-characteristic (ROC) analyses of cross-validation experiments, were measured for FDG PET images obtained with either a 15-mm FWHM PETT V or a 6-mm FWHM Scanditronix PC-1024-7B camera for various methods of data representation. Neural networks were trained to distinguish between normal and abnormal subjects on the basis of regional metabolic patterns. For both databases, classification performance could be improved by increasing the "resolution" of the representation (decreasing the region size) and by normalizing the regional metabolic values to the value of a reference region (occipital region). Results: The optimal classification performance for Scanditronix data (ROC area = 0.95) was higher than that for PETT V data (ROC area = 0.87). Under Bayesian theory, the classification performance with Scanditronix data corresponded to an ability to change a pre-test probability of disease of 50% to a post-test probability of either 90% for a positive classification or 10% for a negative classification. Conclusion: This classification can be used to either strongly confirm or rule out the presence of abnormalities.