Mass Segmentation Results

All the algorithms have been tested on a set of $120$ mammograms extracted from the MIAS database [169] (see Appendix  for more information about this database). Forty of the mammograms contain a mass (see Table  for a breakdown of the dataset). All abnormal mammograms have been manually segmented by a radiologist and the annotations are used in following sections as ground-truth data.

Table 2.3: Test set of mammograms.

		Fatty	Glandular	Dense	Total
-||--	Circumscribed	$9$	$8$	$3$	$20$
	-||--	Spiculated	$6$	$7$	$7$
	-||--	Normal	$28$	$28$	$24$
-||--	Total	$43$	$43$	$34$	$120$
-||--

The first experiment is related to the capability of the algorithms to distinguish mammograms with and without masses, i.e. the capability of the algorithms to detect masses. Therefore, FROC analysis is used in the evaluation.

The second experiment evaluates the accuracy with which the masses have been detected, and here ROC analysis is used to compare the various approaches. For each algorithm and each mammogram, we compute the value of $A_z$ . Thus, mean and standard deviation values for $A_z$ are given for each algorithm.

Figure: FROC analysis of the algorithms over the set of $120$ mammograms.