Table 4. The quantitative validation of DoGNet trained on [Collman15] cAT dataset and applied to [Weiler14] dataset.

Differences with F1 scores on [Collman15] cAT dataset are shown in parentheses.

Method	# params	F1 Score	Prec.	Recall	AUC	|DiC|
ConvNets
Direct	3392	0.72 (0.03)↑	0.79	0.66	0.88	5.33
FCN	3002	0.64 (-0.07)↓	0.85	0.51	0.84	19.
Unet	622	0.79 (0.06)↑	0.85	0.74	0.97	4.33
DoGNets
Shallow Isotropic	62	0.85 (0.1)↑	0.83	0.88	0.96	3.33
Shallow Anisotropic	107	0.83 (0.08)↑	0.88	0.78	0.94	3.33
Deep Isotropic	140	0.88 (0.15)↑	0.83	0.95	0.93	3.33
Deep Anisotropic	230	0.71 (0.0)	0.80	0.63	0.90	7.33
Manually tuned methods
Nieland 2014 [38]	-	0.64 (0.27)↑	0.66	0.62	0.44	2.
Simhal 2017 [32]	-	0.65 (0.0)	0.81	0.55	0.55	13.