Table 5.

Cross-dataset segmentation performances of U-Nets with different training configurations.

Configurations				UKBB test set (n = 600)			ACDC set (n = 100)			BSCMR-AS set (n = 599)
Image resample	Rotation Aug	Flip Aug	Scale Aug	LV	MYO	RV	LV	MYO	RV	LV	MYO
✓	✓	✓	✓	0.923 (0.041)	0.847 (0.038)	0.878 (0.048)	0.873 (0.101)	0.744 (0.104)	0.750 (0.187)	0.851 (0.113)	0.783 (0.095)
✗	✓	✓	✓	0.916 (0.046)	0.836 (0.041)	0.864 (0.053)	0.811 (0.179)	0.614 (0.186)	0.575 (0.270)	0.798 (0.172)	0.673 (0.162)
✓	✗	✓	✓	0.922 (0.042)	0.848 (0.038)	0.878 (0.050)	0.869 (0.117)	0.733 (0.117)	0.722 (0.210)	0.853 (0.118)	0.784 (0.093)
✓	✓	✗	✓	0.924 (0.041)	0.849 (0.037)	0.881 (0.049)	0.858 (0.115)	0.705 (0.142)	0.681 (0.266)	0.862 (0.110)	0.779 (0.092)
✓	✓	✓	✗	0.921 (0.047)	0.845 (0.039)	0.876 (0.050)	0.785 (0.188)	0.640 (0.187)	0.596 (0.279)	0.834 (0.148)	0.752 (0.125)

All experiments were performed with the standard U-Net architecture: UNet-64. Each U-Net was trained using the same UKBB training set for 200 epochs to save computation. Statistics listed in the table are the means and standard deviation of the Dice scores evaluated on the three sets. Numbers in red are those mean Dice scores below 0.70.