Table 3. Comparison of several variations of DoGNets and several baselines on the [Collman15] dataset.

The ‘Shallow3D’ network uses the 3D version of DoGNet, while other variants operate on 2D slices independently. Optimal performance was obtained using Shallow DoGNets.

Method	params	F1 Score	Precision	Recall	AUC	|DiC|
ConvNets
Direct	3392	0.69	0.79	0.62	0.88	11.19
FCN	3002	0.71	0.72	0.70	0.79	4.12
Unet	622	0.73	0.73	0.73	0.91	4.26
DoGNets
Shallow Isotropic	62	0.75	0.74	0.76	0.90	4.25
Shallow Anisotropic	107	0.75	0.75	0.76	0.88	4.26
Shallow3D	61	0.68	0.62	0.77	0.65	9.13
Deep Isotropic	140	0.73	0.77	0.71	0.97	4.99
Deep Anisotropic	230	0.71	0.77	0.33	0.87	7.72
Manually tuned methods
Nieland 2014 [38]	-	0.37	0.49	0.32	0.63	16.5
Simhal 2017 [32]	-	0.65	0.52	0.89	0.74	-