TABLE III. Stage-Specific Segmentation Performance.

Methods	The Harbin dataset
	Early	Progressive	Severe
Percentage	73.1%	18.7%	8.2%
Ours	0.764	0.823	0.875
2D U-net	0.508	0.703	0.755
H-DUnet	0.526	0.788	0.803
MPUnet	0.370	0.618	0.766
3D U-net	0.551	0.802	0.820
3D V-net	0.581	0.796	0.824

The best performer under each criterion is in bold. The performance is shown in the form of average dice.