Groupwise Non-Rigid Registration of Medical Images: The Minimum Description Length Approach
Carole Twining, Stephen Marsland, Chris Taylor · 2004
The aim of non-rigid registration as applied to a group of images is to find a ‘meaningful ’ dense spatial correspondence across the set. There are many methods available for finding such a correspondence given a pair of images, but viewing the groupwise case as just successive pairwise is rather naïve. The principled non-rigid registration of groups of images hence requires a fully groupwise objective function. Statistical analysis of the spatial and pixel-value deformations across the set (as defined by the found correspondence), means that these deformations have to be defined with respect to a common spatial and pixel value reference. We show how the optimal groupwise correspondence can be defined using the Minimum Description Length (MDL) principle, where the definition of the spatial and pixel-value reference is also part of the optimisation. We demonstrate the use of such an objective function as applied to non-rigid registration of a set of 2D T1weighted images of the human brain. As regards constructing the optimal reference image, we show that even in the case when substantial portions of the images are missing, the algorithm not only converges to the correct solution, but also allows meaningful integration of image data across the training set, allowing the original image to be reconstructed as the reference image.