TABLE VII:

Results of a fully-supervised 3D UNet trained with our Shadow-AUG and Shadow-DROP mechanisms where the Shadow-DROP layers are deployed at different stages of the segmentation network. Bold values indicate the best result. Underlined values suggest results without statistical significance compared with the top row (p>0.05).

Models	UCLA dataset			NIH dataset
	DSC[%]	ASD[mm]	HD[mm]	DSC[%]	ASD[mm]	HD[mm]
Shadow-DROP at encoder	92.25(2.19)	0.93 (0.29)	5.89(1.93)	89.85 (3.30)	1.26 (0.58)	6.88 (3.00)
Standard dropout at encoder	91.62(2.48)	1.03(0.39)	6.42(2.27)	81.57(7.40)	2.62(1.31)	13.78(5.92)
Shadow-DROP at bottle-neck	92.17(2.32)	0.94(0.32)	5.93 (1.97)	89.61(3.56)	1.30(0.63)	7.20(3.37)
Standard dropout at bottle-neck	92.32 (2.43)	0.93(0.35)	6.17(2.44)	89.32(6.01)	1.32(0.72)	7.78(3.83)
Shadow-DROP at decoder	92.01(2.41)	0.97(0.34)	6.14(2.22)	89.36(4.23)	1.34(0.70)	7.72(3.85)
Standard dropout at decoder	92.29 (2.40)	0.93 (0.33)	5.76 (2.08)	89.21(6.00)	1.41(1.60)	7.63(5.02)
Shadow-DROP at all layers	92.12(2.32)	0.95(0.31)	5.99(2.04)	89.86 (3.40)	1.29 (0.63)	7.88(4.35)
Standard dropout at all layers	91.48(2.44)	1.05(0.38)	6.48(2.33)	82.75(6.73)	2.27(1.22)	11.82(4.94)