Skip to main content
. 2010 Jul 20;37(8):4155–4172. doi: 10.1118/1.3455704

Figure 8.

Figure 8

Results for the MR-based methods, highlighting classifier regularization trends due to the use of unlabeled data. The difference in AUC [ΔAUC=AUC(with UL data)−AUC(without UL data)] organized according to a quartile decomposition of the initial AUC performance without the use of unlabeled data (lower 25% in blue, lower 25%–50% in light blue, upper 50%–25% in orange, and upper 25% in dark red). In each plot, the quartile dependent change in AUC is ordered according to the use of 50, 100, and 150 L data moving left to right, during training. And within each subset group, the triplet represents the use of a low (50, 100, and 150 UL), medium (400∕500 UL), and high (900 UL) number of UL data. Statistically significant differences from ΔAUC=0 using a paired, nonparametric Wilcoxon signed-rank test, with consideration for multiple-hypothesis testing by employing the Holm–Sidak correction, are indicated by the * above the bars (setting α=0.05 or for adjusted p-values<0.05). The plots display (a) LapSVM for cross-validation, (b) LapSVM with the independent test, and (c) T-LapSVM on the independent test.