As model-based segmentation we include those methods that firstly train the system to detect specific objects. Subsequently, the system has to be able to detect and classify new images depending on the presence or absence of the object. Model-based segmentation methods are closely related to general pattern recognition problems such as pattern matching. In pattern matching, the training is based on images containing the object to detect. Pattern matching has been used in segmentation of mammographic images by Lai et al. [98] and by Constantinidis et al. [35,36]. The main drawbacks of these approaches are the large variation in the shapes of masses and the use of a synthetic model to find real masses. Recently, Tourassi et al. have improved the performance of such approaches by using mutual information as a similarity metric to determine if a (real) query RoI depicts a true mass [179].
Most of the mass segmentation model-based methods train the system on gradient information. Due to the training step we do not classify such approaches as being pure edge-based methods. Training images cover examples with and without the object present. Thus:
Other approaches using edge information are by Jiang et al. [81] and Polakowski et al. [139]. The former was based on the enhancement of the spicules using morphological operations and, subsequently, two features representing the concentration of spicules were used to train a classifier based on a discrimination function. In the work of Polakowski et al. [139], the edges of the image were extracted by subtracting two smoothed versions of the original mammogram. Two Gaussian filters with different standard deviation were used.
On the other hand, Zwiggelaar et al. [203] and Li et al. [103] used statistical approaches to model masses. The former detected spiculated lesions by the union of two techniques: the first one models the centre of the mass using a directional recursive median filter, while the second one models the surrounding pattern of linear structures applying a multi-scale directional line detector. The combination of both methods results in a probability image. Therefore, the detection is performed by thresholding these resulting probability images. On the other hand, Li et al. applied firstly an image enhancement algorithm using morphological filtering. Subsequently, they employed a finite generalized Gaussian mixture (FGGM) distribution to model the histogram. They incorporated an EM algorithm to determine the optimal number of image regions and the kernel shape in the FGGM model. The final step was the use of a Bayesian relaxation labeling approach to perform the selection of suspicious masses.
Three recent approaches [6,24,31] are based on using neural network classifiers and formulate the problem of segmentation as a classification of RoIs being suspicious or not. The features used for the training step are a set of known RoIs containing masses and a set of random samples from normal tissue.