A fast convex optimization approach to segmenting 3D scar tissue from delayed-enhancement cardiac MR images

Abstract

We propose a novel multi-region segmentation approach through a partially-ordered Potts (POP) model to segment myocardial scar tissue solely from 3D cardiac delayed-enhancement MR images (DE-MRI). The algorithm makes use of prior knowledge of anatomical spatial consistency and employs customized label ordering to constrain the segmentation without prior knowledge of geometric representation. The proposed method eliminates the need for regional constraint segmentations, thus reduces processing time and potential sources of error. We solve the proposed optimization problem by means of convex relaxation and introduce its duality: the hierarchical continuous max-flow (HMF) model, which amounts to an efficient numerical solver to the resulting convex optimization problem. Experiments are performed over ten DE-MRI data sets. The results are compared to a FWHM (full-width at half-maximum) method and the inter- and intra-operator variabilities assessed.