Joint Model-Pixel Segmentation with Pose-Invariant Deformable Graph-Priors

This paper proposes a novel framework for image segmentation through a unified model-based and pixel-driven integrated graphical model. Prior knowledge is expressed through the deformation of a discrete model that consists of decomposing the shape of interest into a set of higher order cliques (triplets). Such decomposition allows the introduction of region-driven image statistics as well as pose-invariant (i. e. translation, rotation and scale) constraints whose accumulation introduces global deformation constraints on the model. Regional triangles are associated with pixels labeling which aims to create consistency between the model and the image space. The proposed formulation is pose-invariant, can integrate regional statistics in a natural and efficient manner while being able to produce solutions unobserved during training. The challenging problem of tagged cardiac MR image segmentation is used to demonstrate the performance potentials of the method.