Table 2.

Comparison of different data augmentation (DA) strategies for full lung mask segmentation on ¹H MRI test set (n = 14).

Models	DSC	P value (vs. non-DA model)	MSD (mm)	P value (vs. non-DA model)
Non-DA model	0.955 ± 0.012 (0.947, 0.953, 0.964)	–	1.055 ± 0.953 (0.632, 0.775, 0.916)	–
Conventional-DA model	0.961 ± 0.011 (0.955, 0.963, 0.968)	0.03	0.594 ± 0.137 (0.536, 0.601, 0.680)	0.03
GAN-DA model	0.959 ± 0.011 (0.949, 0.959, 0.967)	0.13	0.928 ± 0.690 (0.550, 0.726, 0.883)	0.31
Combined-DA model	0.965 ± 0.010* (0.955, 0.967, 0.971)	0.0007	0.657 ± 0.254* (0.477, 0.630, 0.715)	0.005

Note. —Data are presented as mean ± standard deviation (25th percentile, median, 75th percentile). DA = data augmentation, GAN = generative adversarial network, DSC = Dice Similarity Coefficient, MSD = mean surface distance. Wilcoxon signed-rank tests with the Bonferroni-Holm correction were used to account for multiple comparisons. Adjusted P-values are reported.

^*Combined-DA model has statistically significantly higher DSC (P = 0.0007) and lower MSD (P = 0.01) compared with GAN-DA model.

DSCs are not statistically significantly different between conventional-DA and GAN-DA, or between conventional-DA and combined-DA.

MSDs are not statistically significantly different between conventional-DA and GAN-DA, or between conventional-DA and combined-DA.