Neural network classification and analysis of positron emission tomography images of subjects with memory disorders

Artificial-neural-network methods were applied to the study of Positron Emission Tomography (PET) images of the brain obtained from subjects studied at two major memory-disorder clinics. The studies involved subjects diagnosed with either "Probable" or "Possible" Alzheimer's Disease, and normal controls. Back-propagation neural networks were trained to distinguish between normal and abnormal subjects on the basis of regional metabolic patterns, as determined by region-of-interest analyses of PET images. A novel, systematic approach was developed to extract and rank the influence of the most-important discriminating features learned by neural networks. This approach required the use of an activation function for neural-network processing units which was "odd" in the mathematical sense, i.e. f(x) = ;The feature-extraction method was used to obtain optimized multivariate discriminating profiles for a variety of training circumstances and forms of data representation. Important aspects of these profiles at the level of brain lobes were: general left-side-low asymmetry, deficits in parietal and temporal regions and preservation of occipital metabolism. On the basis of discriminating profiles for a small-region representation, it was determined that the most valuable regions for purposes of discriminating between normal and abnormal subjects were motor-sensory and mid-temporal regions.