Intensity normalization in the analysis of functional DaTSCAN SPECT images: The α-stable distribution-based normalization method vs other approaches

Abstract

This work is focused on the study of the different intensity normalization procedures available for the proper normalization of 3D functional brain images devoted to the early diagnosis of Parkinson׳s disease. This normalization step is essential, as it corresponds to the initial step in any subsequent computer-based analysis. There are particular features of the Parkinson׳s disease patterns in the functional brain images. Although there are a variety of normalization methods available, these particular features provoke that, for Parkinson׳s disease brain image study, not all of the available normalization procedures present the same proper performance when they are applied. In this work, conventional intensity normalization approaches are considered, along with a novel normalization approach based on the α-stable distribution. All these normalization procedures are referred and their performance is compared. The experimentation and the evaluation results are both based on single-photon emission computed tomography (SPECT) brain images from real cases already diagnosed by expert clinicians. For the performance evaluation, two methods are considered: one based on nearest neighbors and the other related to exceeding a particular striatum activation threshold. In addition, the obtained results are validated by means of statistical analysis, applying the Kullback–Leibler divergence, the Euclidean distance and the Hellinger distance.