Automated Intraventricular Septum Segmentation Using Non-local Spatio-temporal Priors

Automated robust segmentation of intra-ventricular septum (IVS) from B-mode echocardiographic images is an enabler for early quantification of cardiac disease. Segmentation of septum from ultrasound images is very challenging due to variations in intensity/contrast in and around the septum, speckle noise and non-rigid shape variations of the septum boundary. In this work, we effectively address these challenges using an approach that merges novel computer vision ideas with physiological markers present in cardiac scans. Specifically, we contribute towards the following: (1) A novel 1-D active contour segmentation approach that utilizes non-local (NL) temporal cues, (2) Robust initialization of the active contour framework, based on NL-means de-noising, and MRF based clustering that incorporates physiological cues. We validate our claims using cardiac measurement results on approximately 30 cardiac scan videos (approximately 2000 ultrasound frames in total). Our method is fully automatic and near real time (0.1 sec/frame) implementation.