Landmark localization, feature matching and biomarker discovery from magnetic resonance images

The work presented in this thesis proposes several methods that can be roughly divided into three different categories: I) landmark localization in medical images, II) feature matching for image registration, and III) biomarker discovery in neuroimaging. The first part deals with the identification of anatomical landmarks. The motivation stems from the fact that the manual identification and labeling of these landmarks is very time consuming and prone to observer errors, especially when large datasets must be analyzed. In this thesis we present three methods to tackle this challenge: A landmark descriptor based on local self-similarities (SS), a subspace building framework based on manifold learning and a sparse coding landmark descriptor based on data-specific learned dictionary basis. The second part of this thesis deals with finding matching features between a pair of images. These matches can be used to perform a registration between them. Registration is a powerful tool that allows mapping images in a common space in order to aid in their analysis. Accurate registration can be challenging to achieve using intensity based registration algorithms. Here, a framework is proposed for learning correspondences in pairs of images by matching SS features and random sample and consensus (RANSAC) is employed as a robust model estimator to learn a deformation model based on feature matches. Finally, the third part of the thesis deals with biomarker discovery using machine learning. In this section a framework for feature extraction from learned low-dimensional subspaces that represent inter-subject variability is proposed. The manifold subspace is built using data-driven regions of interest (ROI). These regions are learned via sparse regression, with stability selection. Also, probabilistic distributionmodels for different stages in the disease trajectory are estimated for different class populations in the low-dimensional manifold and used to construct a probabilistic scoring function.