Fully automatic segmentation of the open mitral leaflets in 3D transesophageal echocardiographic images using multi-atlas label fusion and deformable medial modeling

The goal of this work is to develop a fully automatic method for segmentation of the mitral leaflets in 3D transesophageal echocardiographic (3D TEE) images. The method combines complementary probabilistic segmentation and geometric modeling techniques to generate 3D patient-specific reconstructions of the mitral leaflets and annulus from 3D TEE image data with no user interaction. In the model-based segmentation framework, mitral leaflet geometry is described with 3D continuous medial representation (cm-rep). To capture leaflet geometry in a target 3D TEE image, a pre-defined cm-rep template of the mitral leaflets is deformed such that the negative log of a Bayesian posterior probability is minimized. The likelihood of the objective function is given by a probabilistic segmentation of the mitral leaflets generated by multi-atlas joint label fusion, while the validity constraints and regularization terms imposed by cm-rep act as shape priors that preserve leaflet topology and constrain model fitting. The method is tested on ten 3D TEE images of human mitral leaflets at mid-diastole, using manual segmentation as the gold standard.