Constraining Contour Deformations Using Statistical A Priori Knowledge of Shape Without Requiring Point-to-Point Correspondence

In this paper we present a method for constraining the deformations of active contour models during image segmentation in a way that is consistent with a previously delineated training set of example images. The training examples are carefully traced with complete or semimanual supervision without the need for pointto-point correspondence. Then frequency domain shape descriptors (in particular, Discrete Cosine Transform coefficients) are used to establish model parameter correspondence. The main modes of shape parameter variation are then captured (using Principal Component Analysis) and used to restrict the active contour model deformations. The method was applied to segment the human left ventricle in echocardiographic images.