Evaluation of different statistical shape models for segmentation of the left ventricular endocardium from magnetic resonance images

Statistical shape models (SSMs) represent a powerful tool used in patient-specific modeling to segment medical images because they incorporate a-priori knowledge that guide the model during deformation. Our aim was to evaluate segmentation accuracy in terms of left ventricular (LV) volumes obtained using four different SSMs versus manual gold standard tracing on cardiac magnetic resonance (CMR) images. A database of 3D echocardiographic (3DE) LV surfaces obtained in 435 patients was used to generate four different SSMs, based on cardiac phase selection. Each model was scaled and deformed to detect LV endocardial contours in the end-diastolic (ED) and end-systolic (ES) frames of a CMR short-axis (SAX) stack for 15 patients with normal LV function. Linear correlation and Bland-Altman analyses versus gold-standard showed in all cases high correlation (r2>0.95), non-significant biases and narrow limits of agreement.