Tissue and Label Modelling for Segmentation of Scar with Contour Correction in Cardiac DE-CMR Volumes

Most scar segmentation methods for cardiac DE-CMR images are constrained by a previous myocardial segmentation that provides borders for the scar identification. However, the automation of these methods rely on an existing myocardial segmentation from CINE-CMR that is registered to the DE-CMR volume; in this step, however, alignment errors are usually introduced and they carry over to the labeling operation. These errors typically remain unchanged after tissue labeling, so inconsistencies between the two (alignment and labeling) may exist. We explore this issue and present a method that, with the same inputs, identifies the healthy and scarred tissue and selectively corrects the endocardial and epicardial contours depending on the image edges, the estimated probabilistic distributions and the proximity to the aligned myocardial borders. For this, we model the posterior probability of each ROI label with a Bayesian approach that unifies the prior tissue probabilities and the myocardial labels. The maximum a posteriori criterion is used to compute a first DE-CMR label map, which is afterwards refined by a connected component analysis. Preliminary results show better accuracy for the endocardial and epicardial contours, and the segmented scar compares favorably with respect to state of the art methods.