MLO Views

We clustered the $35$ MLO views into four different groups according to their mass size. Each group corresponds to the following intervals for mass sizes: $<0.70 cm^2$ , $(0.70-1.20)cm^2$ , $(1.20-2.00) cm^2$ , $(2.00-3.80) cm^2$ . In each interval there were, respectively, $8$ , $8$ , $7$ and $12$ masses. The intervals are different compared to the MIAS database because the masses in this database are smaller.

As already mentioned, the database ground-truth is provided by six experts. Thus, we can compute the performance of the algorithm using those regions where different number of radiologists agree. Considering a true mass those pixels where the six radiologists coincide we obtained $A_z = 90.7\pm 5.8$ . In contrast, if we consider a mass those pixels were at least five radiologists agree, $A_z$ was $90.3\pm 6.2$ . And decreasing the number of agreement we obtained $89.9\pm 6.7$ , $89.6\pm 6.8$ , $89.2\pm 6.9$ , and $88.6\pm 7.0$ for $3$ , $2$ , and $1$ , respectively. This shows an overall trend of performance decrease as the number of radiologist agreement also decreases. This is due to the fact that a different number of thin spicules appears when considering all radiologist annotations as ground-truth. In contrast, only the centre of the mass and clear spicules are taken into account when considering a mass those pixels where all radiologists coincide. In the rest of the evaluation with this database, only this case is analyzed.

The mean $A_z$ for all mammograms for d1 and d2 algorithms was respectively $90.6\pm 8.8$ and $90.6\pm 4.7$ , while our approach obtained $90.7\pm 5.8$ . Note that using this database the overall results obtained by the algorithm d1 are in line with the obtained by the others algorithms. This is due to the fact that the masses in this database are smaller than in the MIAS database, where d1 performs well for small size masses. Table  shows the performance of the algorithms depending on the size of the masses. Note that the same trend shown in the previous section for algorithms d1 and Eig are still valid, and they perform better for smaller masses than for larger ones. In contrast, algorithm d2 shows a similar behaviour for all sizes.

Table 4.2: Influence of the lesion size (in $cm^2$ ) for algorithms d1, d2, and the proposed approach, using the MLO images of Málaga database. The results show mean and standard deviation of the $A_z$ values.

		Lesion Size (in $cm^2$ )
		$<$ 0.70	0.70-1.20	1.20-2.00	$>$ 2.00
-||--	d1	$97.4\pm 3.5$	$88.5\pm 10.5$	$88.4\pm 9.4$	$84.6\pm 10.4$
	-||--	d2	$91.7\pm 3.9$	$89.4\pm 5.1$	$90.5\pm 5.5$
	-||--	Eig	$93.4\pm 4.8$	$91.8\pm 6.6$	$87.7\pm 5.8$
-||--