Spatially Varying Classification with Localization Certainty in Level Set Segmentation

Abstract

We introduce a segmentation framework which extends spatially varying classification to not only incorporate anatomical localization from shape estimation, but to also encode certainty of the localization by local shape variability. The method iterates between a classification step where a statistical classifier learned from feature selection is extended with anatomical localization features, and a shape estimation step where, given the class probability maps, shape is inferred by particle filtering using a level set shape model that accounts for local degrees of anatomical variability. The spatially varying classification is embedded in a geodesic active region framework which allows for local deviations from the inferred shape using an iteratively updated classification based region term. The method is evaluated on late gadolinium enhanced cardiac MRI and is to our knowledge the first automatic segmentation method demonstrated on this type of data.