TABLE 4. Quantitative Comparison of Different Refinement Methods in Stage 2 of 3D Brain Tumor Segmentation.

Method	Dice(%)	ASSD(pixels)
Before refinement	87.55 $\pm$ 6.72	2.04 $\pm$ 1.70
Min-cut user-editing	88.41 $\pm$ 7.05	1.74 $\pm$ 1.53
3D R-Net(Euc)	88.82 $\pm$ 7.68	1.60 $\pm$ 1.56
3D R-Net	89.30 $\pm$ 6.82	1.52 $\pm$ 1.37
3D R-Net(Euc) + CRF-Net(fu)	89.27 $\pm$ 7.32	1.48 $\pm$ 1.22
3D R-Net + CRF-Net(fu)	89.93 $\pm$ 6.49	1.43 $\pm$ 1.16

The segmentation before refinement is obtained by 3D P-Net + CRF-Net(f). 3D R-Net(Euc) uses euclidean distance instead of geodesic distance. Significant improvement from 3D R-Net ($p$-value $<$ 0.05) is shown in bold font.