Classification of positron emission tomography images using multiple kernel learning

Over the last years, several approaches to analyze nuclear medicine imaging using computer systems have been proposed with the aim of assisting the diagnosis of neurodegenerative disorders. Probably one of the most complex challenges facing these approaches is to deal with the huge amount of data provided by brain images. In this work, we propose an original approach based on multiple kernel learning. First the images were parcellated (according to the structure of the brain) by means of the automatic anatomical labeling atlas. Then, the importance of each region for the assisted diagnosis was estimated using a classifi- cation procedure. Finally, all the regions were combined in a multiple kernel method in which one kernel per region was computed and all the kernels were weighted according to the importance of the region they represented. For testing purposes, a database with 46 PET images from stable mild cognitive impairment subjects and early Alzheimer’s disease converter patients was used. An accuracy rate of 73.91% was achieved when differentiating between both groups.