Two-stage active contour model for robust left ventricle segmentation in cardiac MRI

摘要

Segmentation of the endocardial and epicardial boundaries on 3D cardiac magnetic resonance images plays a vital role in the assessment of ejection fraction, wall thickness, end-diastolic volume, end-systolic volume, and stroke volume. Accurate segmentation is significantly challenged by intensity inhomogeneity artifacts, low contrast, and ill-defined organ/region boundaries. We propose a two stage hybrid active contour model for robust left ventricle (LV) segmentation accompanied with a new initialization technique based on prior of the LV structure. The proposed approach includes a new level set method using local, spatially-varying, statistical model for image intensity, an edge-based term to capture region boundaries, and regularization functionals for smooth evolution of the segmenting curve and to avoid expensive reinitialization. Moreover, convex hull interpolation is employed to include the papillary muscles within the endocardial boundary for a refined depiction of LV geometry. The accuracy and robustness of the proposed algorithm were assessed using York, Sunnybrook and ACDC datasets (33 + 45 + 100 subjects), with a wide spectrum of normal hearts, congenital heart diseases, and cardiac dysfunction. Experiments showed that the proposed approach significantly outperformed other active contour methods (overall Dice score 0.90), generating accurate segmentations of left ventricular outflow tract (Dice score 0.91), apical slices (Dice score 0.82), systolic and diastolic phases (Dice scores 0.92 and 0.88 respectively). The percentage of good contours was about 92% and the average perpendicular distance was less than 1.8 mm. Automatically generated segmentation yielded superior estimates of ejection fraction with an R-2 = 0.937. Furthermore, the proposed method was validated using 100 cine MRI cases consisting of five different cardiac classes from the ACDC MICCAI 2017 challenge. The proposed algorithm yielded superior segmentation performance compared with existing active contour models and other state-of-the-art cardiac segmentation techniques, with extensive validation on three standard cardiac datasets, with different cardiac pathologies and phases.