Here we consider a situation where neither gradient-based boundary finding nor region based segmentation gives satisfactory results. Figure 10 (top, left) shows the Magnetic Resonance (MR) image of a dog heart. The aim is to outline the endocardium. The next figure (top, right) shows the probable edge as outlined by an expert. As we can see the image quality is very poor and the edges seem to be very fuzzy. Thus the gradient information is very weak. If we apply gradient based boundary finding, due to a lack of strong edge information the boundary seems to diverge after a few iterations as shown in Figure 10 (bottom, left). Figure 10 (bottom, right) shows the results of the integrated method, which though not perfect, is much better compared to the other method. (It needs to be mentioned, that like all the other experiments, we used the same initialization.) The main reason for this improved performance is that neither region based segmentation nor boundary finding actually fail as there is some information in both the gradient and the region classified image. But by themselves they are not complete and thus neither method produces desirable results. But once we combine them, the output seems reasonable as there is an information fusion, resulting in a better output.
Figure 11 shows a similar sequence of MR brain images, where the task comprises of segmenting the corpus callosum. This is a better quality image compared to the previous application, and thus the improvement in performance is less, but even then one can easily visualize the improvement.
Thus we observe that integration results in a method that is more robust.