3D cardiac segmentation with pose-invariant higher-order MRFs

This paper proposes a novel pose-invariant segmentation approach for left ventricle in 3D CT images. The proposed formulation is modular with respect to the image support (i.e. landmarks, edges and regional statistics). The prior is represented as a third-order Markov Random Field (MRF) where triplets of points result to a low-rank statistical prior while inheriting invariance to global transformations. The ventricle surface is determined through triangulation where image discontinuities can be easily evaluated and the Divergence theorem provides an exact calculation of regional statistics acting on the image or a derived feature space. Promising results using boosting along with the learned prior demonstrate the potential of our method.