Automatic ROI selection using SOM modelling in structural brain MRI

Abstract

This paper presents an automatic method for selecting Regions of Interest (ROI) related to the Alzheimer's Disease (AD) using the information contained in 3D structural MRIs. Normal and AD images are modelled by a number of prototypes provided by Vector Quantization (VQ) algorithms (specifically, Self-Organizing Maps, SOM) which model the volumetric probability histogram and describe the intensity profile of the image. The receptive field of each SOM unit represent a different region of interest on the brain associated to the peaks on the probability histogram. Thus, the space is quantized and the activation level of each SOM unit is associated to the probability of occurrence of the modelled gray level. Additionally, this method can be used to extract a reduced and discriminative features for AD classification, as it compress the information contained in the brain in a reduced number of models. The proposed method has been assessed using the computed ROIs to classify a set of images from the Alzheimer's disease Neuroimaging Initiative (ADNI). © 2013 Springer-Verlag.