Dimensionality Reduction of Medical Image Descriptors for Multimodal Image Registration

Defining similarity forms a challenging and relevant research topic in multimodal image registration. The frequently used mutual information disregards contextual information, which is shared across modalities. A recent popular approach, called modality independent neigh-bourhood descriptor, is based on local self-similarities of image patches and is therefore able to capture spatial information. This image descriptor generates vectorial representations, i.e. it is multidimensional, which results in a disadvantage in terms of computation time. In this work, we present a problem-adapted solution for dimensionality reduction, by using principal component analysis and Horn’s parallel analysis. Furthermore, the influence of dimensionality reduction in global rigid image registration is investigated. It is shown that the registration results obtained from the reduced descriptor have the same high quality in comparison to those found for the original descriptor.