Evaluation of Mass Segmentation Methods

In Section , mass segmentation approaches using one single view have been qualitatively analyzed and classified as region, contour, clustering, or model-based methods. However, based only on this analysis, we can not estimate the robustness of the algorithms with respect to different mammographic cases (different breast tissue, different lesion types, etc.). Therefore, in order to study how the performance varies for the different strategies, and with the aim to extract reliable conclusions, we have quantitatively compared the different approaches. We have selected and implemented two of the most representative algorithms of each class (see Table ).

The algorithms have been evaluated on a set of $120$ mammograms from the MIAS database [169]. Of this set, $40$ mammograms contained masses, whilst the other $80$ represented normal cases. A comparison of the performance between all the algorithms is provided to determine their main advantages and constraints. The performance has been analyzed using ROC and FROC curves [120].

Table 2.2: Compared mass segmentation methods.

Strategy	Id	Method	Based on
Region	$a1$	Region Growing	Petrick et al. [134]
	$a2$	Region Growing	Kupinski and Giger [95]
Contour	$b1$	Laplacian	Petrick et al. [135,136]
	$b2$	Skeletons	Kobatake and Yoshinaga [91]
Clustering	$c1$	K-Means	Sahiner et al. [155,157]
	$c2$	Fuzzy C-Means	Velthuizen [185]
Model	$d1$	Pattern Matching	Lai et al. [98]
	$d2$	Contour + Classifier	Karssemeijer et al. [84]