Interpolation Snakes with Shape Prior for Border Detection in Ultrasound Images

We design interpolation snakes (or I-Snakes ) for border detection in ultrasound images. We argue that interpolation snakes are better suited for ultrasound than other snakes because the control points are on the curve. This allows for arclength parameterization of the snake, and for spatial hard constraints by fixing one or more control points on the spline. Moreover, built into the snake energy is a shape prior term to prevent excessive deviation from a known shape model computed beforehand from training samples. Tests of our algorithm against three other active contour algorithms on a database of 117 kidney ultrasound images show the superiority of the I-snake in this application. Objective measures (symmetric difference, Hausdorff distance and average Hausdorff distance) show that the I-snake generally produces boundaries closer to ground truth boundaries (i.e., delineated by an expert), than other active contours. Also, in a blind study, medical experts chose the better result of two algorithms displayed randomly and the I-snake boundary was strongly preferred over its strongest competitor, the classical Kass-Witkin-Terzopoulos snake. More interestingly, the I-snake boundary was preferred over the ground truth as well.