Table 2.

Intersection over Union metrics of U-Net-based networks trained on the scanning electron microscopy data set for different model initializations and downscaling factors.

#	Model	Model initialization	Downscaling factor	IoUbg	IoUfg
5	Vanilla U-Net	Random	Native	69.2 ± 2.1	77.5 ± 2.4
6	Vanilla U-Net	Random	0.5 × 0.5	75.4 ± 2.9	77.9 ± 3.3
6v	Vanilla U-Net	Random	0.5 × 0. 5*	72.7 ± 2.3	79.9 ± 2.8
7	U-Net VGG16	Pre-trained	Native	71.0 ± 2.5	80.1 ± 2.4
8	U-Net VGG16	Pre-trained	0.5 × 0.5	77.7 ± 3.2	80.4 ± 2.6
8v	U-Net VGG16	Random	0.5 × 0.5	68.4 ± 4.0	73.1 ± 1.5

The superscript ^v indicates a validation experiment conducted to test a particular hypothesis, as described in the section “Segmentation results”. The superscript * indicates that downscaling was performed after tiling.