Novel Methods for Multi-Shape Analysis

Multi-shape analysis has the objective to recognise, classify, or quantify morphological patterns or regularities within a set of shapes of a particular object class in order to better understand the object class of interest. One important aspect of multi-shape analysis are Statistical Shape Models (SSMs), where a collection of shapes is analysed and modelled within a statistical framework. SSMs can be used as (statistical) prior that describes which shapes are more likely and which shapes are less likely to be plausible instances of the object class of interest. Assuming that the object class of interest is known, such a prior can for example be used in order to reconstruct a three-dimensional surface from only a few known surface points. One relevant application of this surface reconstruction is 3D image segmentation in medical imaging, where the anatomical structure of interest is known a-priori and the surface points are obtained (either automatically or manually) from images. Frequently, Point Distribution Models (PDMs) are used to represent the distribution of shapes, where each shape is discretised and represented as labelled point set. With that, a shape can be interpreted as an element of a vector space, the so-called shape space, and the shape distribution in shape space can be estimated from a collection of given shape samples. One crucial aspect for the creation of PDMs that is tackled in this thesis is how to establish (bijective) correspondences across the collection of training shapes. Evaluated on brain shapes, the proposed method results in an improved model quality compared to existing approaches whilst at the same time being superior with respect to runtime. The second aspect considered in this work is how to learn a low-dimensional subspace of the shape space that is close to the training shapes, where all factors spanning this subspace have local support. Compared to previous work, the proposed method models the local support regions implicitly, such that no initialisation of the size and location of these regions is necessary, which is advantageous in scenarios where this information is not available. The third topic covered in this thesis is how to use an SSM in order to reconstruct a surface from only few surface points. By using a Gaussian Mixture Model (GMM) with anisotropic covariance matrices, which are oriented according to the surface normals, a more surface-oriented fitting is achieved compared to a purely point-based fitting when using the common Iterative Closest Point (ICP) algorithm. In comparison to ICP we find that the GMM-based approach gives superior accuracy and robustness on sparse data. Furthermore, this work covers the transformation synchronisation method, which is a procedure for removing noise that iii Acknowledgements accounts for transitive inconsistency in the set of pairwise linear transformations. One interesting application of this methodology that is relevant in the context of multi-shape analysis is to solve the multi-alignment problem in an unbiased/reference-free manner. Moreover, by introducing an improvement of the numerical stability, the methodology can be used to solve the (affine) multi-image registration problem from pairwise registrations. Compared to reference-based multi-image registration, the proposed approach leads to an improved registration accuracy and is unbiased/reference-free, which makes it ideal for statistical analyses. iv