Multivariate analysis of spect images with illustrations in Alzheimer's disease

Modern neuroimaging techniques such as positron emission tomography and single photon emission computed tomography (SPECT) often generate datasets consisting of a large number of variables measured on a small number of subjects. In addition, the data contain large 'subject effects' which must be adjusted for in any statistical analysis. This paper illustrates the dangers inherent in naive univariate analysis and proceeds to demonstrate the application of multivariate methods such as Hotellings T2-test, canonical correlation analysis and discriminant analysis. The issues and statistical methods are illustrated using a dataset from a SPECT study of groups of normal control subjects and subjects with a clinical diagnosis of Alzheimer's disease.