Skip to Main Content
Expert physicians are able to attain good Alzheimer's Disease (AD) diagnostic accuracy, relying on visual inspection of Positron Emission Tomography (PET) images only. Nevertheless, computerized methods have been implemented with similar or even better performance. We investigate the potential of the physician's experienced visual inspection to guide feature selection, in an automatic classification procedure. Eye tracking methodology is employed to obtain a model of the physician's visual behavior, which allows for the sampling of voxel intensity features that are then fed to an SVM classifier. This approach is compared with commonly used automatic feature selection alternatives. Image data were taken from the Alzheimer's Disease Neuroimaging Initiative database. The results show that the proposed approach marginally improves accuracy in AD vs. CN classification, but for MCI vs. CN and AD vs. MCI it presents lower performance.