We have presented in this paper a new technique for integrating the method of region based segmentation into gradient based boundary finding. This is separate from the other related works in the same way as edge detection is different from boundary finding. Our method is based on the Bayesian optimization theory of maximization of the a posteriori probability. As the examples show, the integrated approach is much more robust to both increased amounts of noise as well as increasingly displaced initialization of the initial boundary. Almost uniformly there is an improvement over the conventional gradient based boundary finding. To prove this, we have devised a variety of experiments and the results from all of them look favorable.
Application of this method on real medical images result in noticeable improvent as has been shown. We have started using it for clinical research purposes recently for outlining the endo and epicardium of heart images and the results are much better than what we had acheived the method reported in [1].
Even though this whole work is a 2-D effort, conceptually, it can be carried over to 3-D without any trouble. Our next effort will be in that direction. Also, it would be interesting to see how this process of refinement of one method using related output of another method can be used in a recursive mode until we come to some sort of an equilibrium. What we have in our mind is a game theoretic framework [13], [14], where the players correspond to the two methods of region based segmentation and boundary finding.