Medial-guided fuzzy segmentation

Abstract

Segmentation is generally regarded as partitioning space at the boundary of an object so as to represent the object's shape, pose, size, and topology. Some images, however, contain so much noise that distinct boundaries are not forthcoming even after the object has been identified. We have used statistical methods based on medial features in Real Time 3D echocardiography to locate the left ventricular axis, even though the precise boundaries of the ventricle are simply not visible in the data. We then produce a fuzzy labeling of ventricular voxels to represent the shape of the ventricle without any explicit boundary. The fuzzy segmentation permits calculation of total ventricular volume as well as determination of local boundary equivalencies, both of which are validated against manual tracings on 155 left ventricles. The method uses a medial-based compartmentalization of the object that is generalizable to any shape.