Segmentation of the Plaque

The segmentation of the plaque is a really important step before tissue characterization. There are multiple ways to achieve this goal [19,20,43-45]. In particular, we will focus on two general lines of work: the line proposed in [43] and the line proposed in [19,20].

In [43] the process of segmentation relies on a manual definition of a region of interest. Using that region of interest a Sobel-like edge operator with neighborhoods of 5 x 5 and 7 x 7 is applied. Once we have these features extracted, the problem of identification of the vessel wall and plaque border is solved by finding the optimal path in a two-dimensional graph. The key of the graph searching is to find the appropriate cost functions. In the paper, the authors propose different cost functions depending on whether the lumen-plaque border or the adventitia border is desired.

Having in mind the tissue classification goal of the process, in [19,20], the authors try to find a segmentation of the overall tissue independent of what kind of tissue it is, to distinguish the lumen-plaque border. Therefore, the method consists of selecting a feature space and a classifier. This method takes advantage of the fact that for tissue characterization the same scheme must be used. Thus, a classifier is trained for general tissue discrimination. Hence, in the overall process the feature extraction process is performed once for both, plaque segmentation and tissue identification. What is different in both approaches is the classification selection and training data, and the post processing steps.

The classification step is performed using a fast classifier, boosting methods, or ML. The result of this step is a series of unconnected areas that are related to tissues. In order to find the exact location of the lumen-plaque border, a fast parametric snake is let to deform over the unconnected areas [46]. The snake performs a double task: first, it finds a continuous boundary between blood and plaque. The second task is that it ensures an interpolation and a fill-in process in regions where tissue is not located or not reliable (such as areas with reverberations due to the guide-wire, etc). The adventitia border is found by context using a 5 x 5 Sobel-like operator and deformable models. Figure 2.19 shows an example of a possible scheme for border location. First the IVUS image is transformed to cartesian coordinates. Then, a texture feature extraction step is performed. A classification scheme is trained to distinguish blood from tissue. At the end, a snake deforms to adapt to the classified IVUS image and locates accurately the blood-plaque border.

0 0

Post a comment