Interpolation snakes with shape prior for boundary detection in noisy images

We develop an active contour algorithm (I-SNAKE) designed to cope well with the noise, specularity and incomplete visual support typically found in ultrasound images. To impose smoothness, we model mathematically our active contour as an interpolation spline. This model also allows a user to set hard spatial constraints on the snake by fixing one or more control points of the spline. We introduce a shape prior term in the snake energy functional that acts as shape memory and prevents excessive deviations from a known shape model computed from shape training samples. This shape model also supports operation with occlusion and loss of contrast. Extensive tests on 117 ultrasound images comparing the similarity of the detected boundaries to manually traced contours (ground truth) showed that the interpolation snake algorithm (I-SNAKE) did significantly better than three other active contour algorithms: the original Kass-Witkin-Terzopoulos snakes (KWT), the Geodesic Active Contour model (GAC) and the Chan-Vese model (CV). Moreover, for clinical validity a subjective evaluation by a human expert was performed via a blind study that compared I-SNAKE with the ground truth (GT) or with the KWT model. The I-SNAKE boundary was preferred in 80 out of 117 comparisons with GT, and in 99 out of 117 with KWT. Other example results are given in several different image domains.