Constrained active appearance models for segmentation of triplane echocardiograms

This paper presents multiview and multiframe active appearance models (AAMs) for left ventricular segmentation in triplane echocardiograms. We describe a general way of integrating local edge detector based segmentation algorithms into the AAM framework. The feasibility of this approach is evaluated by comparing an AAM constrained by a dynamic programming (DP) based snake with an unconstrained AAM, and an AAM constrained by manually defined landmarks. A leave-one-out validation scheme was used for training and testing of the methods. Evaluation was done in 36 patients suffering from various heart diseases, using manually determined volumes and ejection fractions (EF) as reference. The segmentation was initialized by manual selection of the mitral annulus and apex in three imaging planes. The differences, in volume, between manual segmentation and the best automatic method (DP-constrained AAM) were -3.1 +/- 20 ml (mean +/-SD) at end-diastole and 0.61 +/- 13 ml at end-systole. The difference in EF was -1.3 +/- 6.3%, comparable to the interobserver variability. We show that 1) constraining the model to manually defined landmarks improves volume and EF estimates compared to unconstrained AAMs, 2) further improvement is achieved using a DP-constrained AAM, and 3) segmentation in triplane echocardiograms gives higher accuracy than single plane data.