PARAMETRIZATION OF LEVEL-SETS WITH B-SPLINES

Level-sets are powerful techniques to segment images because they can accommodate any contour topologies. We used B-splines to model level-set functions using fewer knots/coefficients than pixels. This forces the contours to be smooth without the need to minimize smoothing terms. We implemented a standard variational method where objects were segmented based on their edges. We also developed a method to segment images of piecewise constant intensity objects. In this case the level-sets were directly computed from a classification step without evolving the contours. We tested our method on simulated MRI brain data. We showed that by using three level-sets in a multi-layer scheme, the classification of brain tissues was more robust than the standard fuzzy c-means algorithm even with spatial regularization.