A Brief Review on Deformable Template Models

When talking about deformable template models, people usually think of snakes. However, snakes are just a kind of specific template models, as shown in the survey of Jain and Dubes [79], where deformable template models are classified according to Figure . This section briefly describes the different categories explained in this survey.

There are two main trends on template models: those which deal with a rigid (fixed) template and the rest where this template varies. It should be noted that the former trend is composed by the early template matching approaches and, although they can be applied on some industrial applications, nowadays are less used. Despite this fact, the reviewed pattern matching approaches for mass segmentation belong to this class of algorithms.

Figure 4.1: An overview of the template matching techniques. Extracted from the survey of Jain et al. [79].

In contrast a deformable model is active in the sense that it is able to adapt itself to fit the given data. Two different sub-trends can be found in this direction: free-form models and parametric models. The former models can represent any arbitrary shape as long as some regularization constraint (continuity, smoothness) is satisfied. The well known active contours approaches (snakes) [85] are classified in this category. Examples of snakes used in mass segmentation reviewed in Chapter  are the works of Kobatake et al. [89] and Sahiner et al. [156,158]. Both approaches used such technique to refine a previous rough segmentation.

On the other hand, there are approaches that provide information related to the shape of the object to the system. Typically, these works firstly try to characterize the shape using a parametric formula, and secondly define a set of deformation modes which let the initial shape vary and adapt to the real images. As usually it is difficult to find a parametric formula to describe the shape, and therefore a prototype template is also used. The active shape models algorithm [37] is the most common algorithm of this category. In this algorithm, a database of manually segmented images is necessary to construct the mean shape and find their modes of deformation using PCA analysis. In a second step, a new image is segmented using its gradient description and finding which are the main deformed modes. However, an important drawback of such algorithm is the tricky manual segmentation, marking the same number of points at the same position. As is shown in Figure  it is even difficult to say that two masses have similar positions, due to the large variation found in masses.

Figure 4.2: Four RoIs corresponding to manually detected masses.

Thus, according to Figure  the algorithm proposed in this work should be classified as a prototype parametric deformable template matching algorithm. By means of the application of the eigenfaces algorithm [180] over a set of real masses, a template and its deformation's modes are found. Subsequently, and using a Bayesian scheme, the prototype is searched in the images. In contrast to Active Shape Models, the initial database of our proposal can be easily obtained from the different public mammographic databases, as only a rough manual segmentation is needed. Concretely, only the centre and the size of the masses are necessary as a starting point (just the bounding box of the mass).