Table 1.

The comparison of different types of mistakes generated by watershed 2D + Link or by watershed 3D from the EM demo dataset in slice # 1 and 2. The boundary mistakes were manually counted (see Supplementary Fig. S5). The under-segmentation and over-segmentation mistakes were defined as the ratio of cells in ground truth images with mistakes, calculated automatically with a python program using the ground truth data.

Methods	Boundary mistakes		Under-segmentation		Over-segmentation
	Slice #1	#2	#1	#2	#1	#2
Watershed 2D + Link	0	0	45%	46%	42%	35%
Watershed 3D	67	69	99%	97%	71%	72%