Comparison between Parzen window interpolation and generalised partial volume estimation for nonrigid image registration using mutual information Dirk Loeckx, Frederik Maes, Dirk Vandermeulen, and Paul Suetens Medical Image Computing (ESAT/PSI), Faculty of Engineering University Hospital Gasthuisberg, Herestraat 49, B-3000 Leuven, Belgium. Dirk.Loeckx@uz.kuleuven.ac.be Abstract. Because of its robustness and accuracy for a variety of ap- plications, either monomodal or multimodal, mutual information (MI) is a very popular similarity measure for (medical) image registration. Cal- culation of MI is based on the joint histogram of the two images to be registered, expressing the statistical relationship between image inten- sities at corresponding positions. However, the calculation of the joint histogram is not straightforward. The discrete nature of digital images, sampled as well in the intensity as in the spatial domain, impedes the exact calculation of the joint histogram. Moreover, during registration often an intensity will be sought at a non grid position of the floating image. This article compares the robustness and accuracy of two common his- togram estimators in the context of nonrigid multiresolution medical image registration: a Parzen window intensity interpolator (IIP) and generalised partial volume histogram estimation (GPV). Starting from the BrainWeb data and realistic deformation fields obtained from pa- tient images, the experiments show that GPV is more robust, while IIP is more accurate. Using a combined approach, an average registration er- ror of 0.12 mm for intramodal and 0.30 mm for intermodal registration is achieved. 1 Introduction The goal of image registration is to find a transformation that maps positions of a reference image I R onto the corresponding positions of a floating image I F and is optimal in some sense. Different ways exist to judge the similarity between the reference and (deformed) floating image. They can be broadly classified into two categories: feature based and intensity based methods. In 1995, Collignon et al. [1] and Viola et al. [2] independently introduced mutual information (MI) as a similarity measure for intensity based medical image registration. Because of its robustness and accuracy for a variety of applications, either monomodal or multimodal, its popularity has been growing ever since [3, 4].