Table 2.

Best classification performance on BCDR database

		Normal/Abnormal	Benign/Malignant
Embedded Method	AUC	98.16 (97.87−98.48)∗∗	92.08 (91.61−92.58)
	Accuracy	97.31 (96.92−97.31)∗∗	88.46 (87.69−89.23)
	Sensitivity	94.62 (93.85−94.62)	89.09 (87.27−90.91)
	Specificity	100 (100−100)∗∗	88.00 (86.67−89.33)
Filter Method	AUC	98.67 (98.57−98.76)	92.13 (91.66−92.78)
	Accuracy	96.92 (96.54−96.92)	87.69 (86.92−89.23)
	Sensitivity	93.85 (93.85−94.62)	89.09 (87.27−90.91)
	Specificity	99.23 (99.23−100)	87.33 (85.33−89.33)

The classification performance calculated in correspondence with the best result highlighted in the 100 rounds of 10-fold cross-validation for increasing the number of selected features, are summarized. We tested the significance of the diversity of performance measures obtained with the two different feature selection techniques on the same classification problem. Statistical significance is measured with the Wilcoxon-Mann-Whitney test: ** p-value <0.01 (Bonferroni correction)