False Positive Reduction Step with Breast Density Information

In this section we exhaustively evaluate the effect of introducing the breast density into the false positive reduction step. As already explained, this is done by clustering the DDSM RoIs database according to this parameter. The results are obtained by using the same leave-one-out method explained in Section , where $256$ RoIs depicted a true mass and the rest $1536$ were normal, but suspicious tissue.

Figure  shows the mean $A_z$ value obtained using the leave-one-out strategy and varying the ratio between both kinds of RoIs. Note that, as expected, the performance of both PCA and 2DPCA approaches decreases as the ratio of RoIs depicting masses decrease. For the PCA approach we obtained $A_z = 0.81$ for the ratio $1/1$ and $A_z = 0.71$ for the ratio $1/6$ , while using the 2DPCA approach we obtained $A_z = 0.96$ and $A_z = 0.85$ respectively. Again, the 2DPCA approach obtained better performances than the PCA.

Figure: Mean $A_z$ value of the system for the DDSM database at different RoIs ratio.

The mean for each cluster size at ratio $1/3$ is shown in Table . The overall performance of the system is up to $0.91$ . Moreover, a similar trend to the one mentioned in  is observed, with a better classification for larger masses. Thus, comparing the performance of both results we show that considering the breast tissue obtain an improvement of $0.05$ in $A_z$ value.

Table 6.4: $A_z$ results for the classification of masses taking the breast density into account at ratio $1/3$ .

		Lesion Size (in $cm^2$ )
		$<$ 0.10	0.10-0.60	0.60-1.20	1.20-1.90	1.90-2.70	$>$ 2.70
-||---	PCA	$0.70$	$0.71$	$0.71$	$0.72$	$0.77$	$0.89$
	-||---	2DPCA	$0.88$	$0.93$	$0.91$	$0.92$	$0.89$
-||---

Figure 6.4: Performance of the system for the DDSM database.

Figure  shows the mean kappa value obtained using the leave-one-out strategy at different ratios and threshold $0.5$ . The same behaviour mentioned for $A_z$ values is repeated, where the performance of both approaches are reduced when increasing the number of normal samples. On the other hand, Figure  shows a comparison for the 2DPCA approach with and without taking breast density information into account. Note that considering such parameter the results clearly improves. At ratios $1/1$ , $1/2$ , $1/3$ , and $1/4$ the agreement can be considered as almost perfect, while for ratios $1/5$ and $1/6$ the agreement is substantial. In contrast, without using such information, the agreement for the two latter ratios was only moderate, and only the agreement in ratios $1/1$ and $1/2$ could be considered as almost perfect.

Figure 6.5: Comparison of the false positive reduction algorithm without and with using breast density information.

Finally, in Table  a comparison of the performance of the methods with respect to each BIRADS category is shown. Note that for the PCA-based method, mammograms with lower BIRADS were better classified that mammograms with higher BIRADS. This result seems plausible because it is equivalent to the performance of a human expert, as it is well known that experts radiologists have more difficulties to find masses in dense mammograms than in fatty ones. In contrast, the performance of the 2DPCA-based method is more independent of the breast tissue, although for BIRADS IV its performance decreased.

Table 6.5: $A_z$ results for each BIRADS category.

		Breast Tissue
		BIRADS I	BIRADS II	BIRADS III	BIRADS IV
-||--	PCA	$0.83$	$0.91$	$0.80$	$0.75$
	-||--	2DPCA	$0.94$	$0.93$	$0.94$
-||--