An extended strategy for exploratory multivariate image analysis including noise considerations

Abstract

AbstractMultivariate image analysis (MIA) is a powerful tool for many image segmentation and classification problems, but the interpretation and understanding of the original and resulting multidimensional (multivariate) data are not always easy. A strategy for MIA has been proposed which describes its usage on multivariate images for segmentation tasks. MIA starts with principal component analysis (PCA) and then continues with interactive analysis of the output from PCA. In this paper a number of extensions to MIA are proposed. The extensions are the suggestion to incorporate preprocessing of the multivariate image in MIA, the suggestion to use synthetic multivariate image models which create a clear‐cut situation, and new visualization tools which improve the interactivity and understanding of the results. Extended MIA is applied on synthetic multivariate image data simulating a possible application with large noise, positron emission tomography (PET). As a result of the interactive analysis, suggestions for preprocessing emerge. The developed methodology for handling the noise is then applied on real PET image data with good results.