Abstract

Currently, there is no cure for Alzheimer's disease, but its early detection is essential to an effective treatment, slowing down the progression of symptoms. Consequently, the development of automatic diagnostic tools, which use as principal source of information three-dimensional images of the brain, has attracted great interest in recent years. This work focused on PET images and studied alternatives to two of the main building blocks of a computerized diagnostic system: the extraction and selection of features. Regarding the common approach based on Voxel Intensities, the FDG-PET image was studied for different scales and resolutions. In addition, the use of a measure of local contrast was also tested, as well as the widely known texture descriptor, Local Binary Patterns, to which a novel extension to three dimensional data was proposed. As regards selection, a new method based on data acquired by the Eye Track technology during the inspection of PET images by an expert physician was proposed. The aim of this method is to model the behavior of the gaze over time, and use the model to select the features that the expert found most interesting. Moreover, other more conventional methods based on correlation measures and mutual information were also studied. The Support Vector Machine classifier was used to perform binary classifications among AD patients, patients with Mild Cognitive Impairment and a control group (in a dichotomous fashion), obtaining comparable or superior performances to those achieved by most systems found in the literature.

Back