Using automated texture features to determine the probability for masking of a tumor on mammography, but not ultrasound

Table 3 Selected texture features for predicting percentage mammographic density (PMD)

Feature family	Number of features				Correlation with PMD Median (min., max.)^c
Feature family	All	Lasso^a	Boosting^a	Common^b	All	Lasso	Boosting
Fourier	12	9	9	6	0.16 (0.03, 0.28)	0.10 (0.03, 0.28)	0.12 (0.03, 0.28)
Histogram	14	13	13	12	0.18 (0.00, 0.25)	0.19 (0.00, 0.25)	0.17 (0.00, 0.25)
Markovian	37	24	24	20	0.44 (0.00, 0.72)	0.39 (0.00, 0.72)	0.43 (0.00, 0.72)
Moment-based	70	54	33	32	0.21 (0.00, 0.61)	0.21 (0.00, 0.61)	0.21 (0.00, 0.61)
Regional	45	36	32	28	0.22 (0.01, 0.52)	0.23 (0.03, 0.52)	0.20 (0.01, 0.52)
Run length	28	18	15	12	0.59 (0.04, 0.71)	0.60 (0.15, 0.70)	0.62 (0.04, 0.71)
Wavelet	12	10	8	8	0.24 (0.01, 0.42)	0.26 (0.06, 0.42)	0.26 (0.06, 0.31)
Total	218	164	134	118

^aSelected number of features using lasso and boosting method, respectively, to predict PMD. Prediction models were fitted on the complete dataset. The tuning parameters were estimated by cross-validation
^bNumber of features selected both by lasso and boosting
^cEach feature was correlated with PMD. Summary statistics (median, minimum, maximum) of Spearman correlation coefficients between (all and selected) features and PMD are shown

ISSN: 2047-783X