Automated MRI-based biventricular segmentation using 3D narrow-band statistical level-sets

The goal of this study was to develop a near-automated technique for the segmentation of left ventricular (LV) endo- and epicardial as well as right ventricular (RV) endocardial contours from cardiac magnetic resonance (CMR) images. The newly developed technique was tested against conventional manual tracing. Our approach is based on a 3D narrow-band statistical level-set algorithm (applied to a stack of CMR short-axis images) followed by several refinement steps. This technique was tested on steady-state free precession (SSFP) CMR images acquired during 10-15 sec breathholds in 6 patients, including a total of 120 images. Computational time was around 3 min for a stack of 10 slices. For performance evaluation, an experienced interpreter manually traced ventricular contours on all the images. Quantitative error metrics (Hausdorff distance, HD; mean absolute distance, MAD, Dice coefficient, DC) were computed between automatically identified and manually traced contours. Bland-Altman and linear regression analyses were also performed between automatically and manually computed ventricular volumes. The results (MAD: LV Endo = 1.3±0.7 px, RV Endo = 1.7±1.2 px, LV Epi = 1.5±0.7 px) indicate that fast and accurate identification of LV and RV contours using 3D narrow-band statistical level-sets is feasible.