Contour-Based Methods

Image segmentation techniques based on edge detection have been in use since the early work of Roberts [152]. However, identifying regions on the basis of edge information is far from trivial, since algorithms for edge detection do not usually possess the ability of the human vision system to complete interrupted edges using experience and contextual information. Therefore, sometimes edges are detected which are not the transition from one region to another and correctly detected edges often have gaps in them at places where the transitions between regions are not abrupt enough. Hence, detected edges may not necessarily form a set of closed connected curves that surround distinct regions.

As indicated in Table , there is only a limited number of publications trying to segment mammographic images using edge-based methods, which is mainly due to the difficulty of extracting the boundary between masses and normal tissue. Typical algorithms for finding edges are based on filtering the image in order to enhance relevant edges prior to the detection stage. The earliest approaches for mass segmentation are based on such methodology. The detection of edges in Petrick et al. [135,136] was based on a Gaussian-Laplacian edge detector, after the image was enhanced by the adaptive density-weighted contrast enhancement filter. A different approach is described by Kobatake and Yoshinaga [91], which starts with a sub-image containing a possible mass lesion. The algorithm looks for spicules using gradient information in three steps: firstly, the morphological line-skeletons are extracted in order to detect long and thin anatomical structures (like spicules). Secondly, a modified Hough transform is used to extract lines passing near the centre of the mass, and finally the algorithm automatically selects candidates based on the number of line-skeletons that satisfy the second step.

In recent approaches, edge information has been used to refine initial segmentation results. Examples are Kobatake et al. [89] and Sahiner et al. [156,158], who used active contour models (snakes) as a final step of their algorithms. Timp and Karssemeijer [177] found the best contour of the mass by an optimization technique based on dynamic programming. Their approach used both edge based information as well as a priori knowledge about the grey-level distribution of the region of interest around the mass. They demonstrated a better performance of their method in comparison with an implemented version of the region growing algorithm inspired on the already mentioned work of Kupinski and Giger [95], and the discrete contour model inspired on the work of Viergever and Lobregt [107].

There are related approaches, which are based on the detection of spicules and the use of statistical analysis of gradient-orientation maps [84,175]. However, due to the necessity to perform a posterior classification step we consider these as model-based approaches.