TABLE 2.

Ablation study performed by comparing the segmentation accuracy between different models and their corresponding 3D‐MASNet in terms of DICE by two‐fold cross validation

Network	CSF		GM		WM		Avg
	Baseline	MixACB	Baseline	MixACB	Baseline	MixACB	Baseline	MixACB
BuiNet	0.938 ± 0.010	0.938 ± 0.011	0.905 ± 0.007	0.908* ± 0.007	0.888 ± 0.014	0.892* ± 0.013	0.910 ± 0.007	0.912* ± 0.007
3D‐UNet	0.940 ± 0.010	0.942* ± 0.008	0.907 ± 0.007	0.909* ± 0.007	0.889 ± 0.014	0.892* ± 0.015	0.912 ± 0.007	0.914* ± 0.008
CC‐3D‐FCN	0.923 ± 0.010	0.942* ± 0.008	0.910 ± 0.006	0.911 ± 0.007	0.892 ± 0.013	0.894* ± 0.013	0.908 ± 0.006	0.915* ± 0.006
NLU‐Net	0.947 ± 0.009	0.949* ± 0.008	0.918 ± 0.007	0.919 ± 0.006	0.903 ± 0.012	0.904 ± 0.014	0.922 ± 0.006	0.924* ± 0.006
DU‐Net	0.951 ± 0.008	0.953* ± 0.008	0.922 ± 0.007	0.923* ± 0.007	0.907 ± 0.015	0.907 ± 0.015	0.927 ± 0.007	0.928* ± 0.008

Note: The best values are highlighted in bold font. “Baseline” denotes that the corresponding model adopted the standard convolutional operation; “MixACB” denotes that the corresponding model was transformed into 3D‐MASNet; “*” denotes that the difference between baseline and 3D‐MASNet is statistically significant (p < .05) .