Diagnosis and modelling of Alzheimer's disease through neural network analyses of PET studies

Abstract

The back-propagation neural network algorithm was applied to the analysis of regional patterns in cerebral function, as demonstrated in positron emission tomography (PET). A trained network was able to successfully distinguish PET scans of normal subjects from PET scans of Alzheimer's Disease patients. It is concluded that the combination of PET and neural networks is a useful diagnostic tool for Alzheimer's Disease. A new paradigm for back-propagation learning is discussed which emphasizes its similarity to template matching. It is demonstrated that, under certain circumstances, the back-propagation network can be used as an estimation tool, as well as a classification tool, i.e., a trained neural network can indicate the criteria by which its classifications are performed.