Principal component analysis of dynamic PET and gamma camera images: a methodology to visualize the signals in the presence of large noise

Abstract

Principal component analysis (PCA) is a data-driven technique used to explain the variance-covariance structure of a data set. PCA of noisy image data can be expected to be hard to perform properly, since PCA has no way to discriminate between variance due to signals and variance due to noise. Further, PCA call not discriminate between pixels belonging to the background and pixels belonging to the object(s). The authors show that PCA of gamma camera and positron emission tomography (PET) images can be significantly improved by taking the noise and spatial background into consideration. The two applications represent two fundamentally different noise problems, namely large background noise and signal dependent noise. The problems are illustrated using a synthetic image and a methodology for exploring the feature space called multivariate image analysis (MIA). After defining the problems, a methodology for handling the noise is proposed. The preprocessing which is proposed is equivalent to expressing pixels according to their significance levels.<>