PET imaging analysis using a parcelation approach and multiple kernel classification

Abstract

Positron Emission Tomography (PET) is a noninvasive medical imaging modality that provides information about physiological processes. Due to its ability to measure the brain metabolism, it is widely used to assist the diagnosis of neurodegenerative disorders such as Alzheimer's disease (AD) of Parkinsonism. In order to avoid the subjectivity inherent to the visual exploration of the images, several computer systems to analyze PET data were developed during the last years. However, dealing with the huge amount of information provided by PET imaging is still a challenge. In this work we present a novel methodology to analyze PET data that improves the automatic differentiation between controls and AD patients. First the images are divided into small regions or parcels, defined either anatomically, geometrically or randomly. Secondly, the accuray of each single region is estimated using a Support Vector Machine (SVM) classifier and a cross-validation approach. Finally, all the regions are assessed together using multiple kernel SVM with a kernel per region. The classifier is built so that the most discriminative regions have more weight in the final decision. This method was evaluated using a PET dataset that contained images from healthy controls and AD patients. The classification results obtained with the proposed approach outperformed two recently reported computer systems based on Principal Component Analysis and Independent Component Analysis.