Table 1.

Data augmentation techniques applied in the approaches validated within the BraTS 2018 challenge framework.

References	Model	Flip	Rot.	Trans.	Scale	Shear	Elastic	GAD	Pixel-wise
Albiol et al., 2019	VGG, Inception, Dense	3D affine transformations
Benson et al., 2018*	CNN (encoder-decoder)	Yes							Random
Carver et al., 2018	U-Net	Yes
Chandra et al., 2018	V-Net, ResNet-18, FC-CRF	Yes			Yes
Dai et al., 2018	Domain-adapted U-Net	Yes
Feng et al., 2018	U-Net	Yes
Gholami et al., 2018*	U-Net							PDE
Isensee et al., 2018	U-Net	Yes	Yes		Yes		Random		Gamma
Kao et al., 2018	DeepMedic, 3D U-Net	Yes
Kermi et al., 2018	U-Net	Yes	Yes	Yes
Lachinov et al., 2018*	Cascaded U-Net	Yes					B-spline		Gaussian
Ma and Yang, 2018	3D CNN	Yes	Yes		Yes
McKinley et al., 2018	Dense CNN	Yes	Yes						Shift, scale
Mehta and Arbel, 2018	U-Net		Yes	Yes	Yes	Yes
Myronenko, 2018	CNN (encoder-decoder)	Yes			Yes				Shift
Nuechterlein and Mehta, 2018	3D-ESPNet	Yes			Yes
Puybareau et al., 2018	VGG-16		Yes		Yes
Rezaei et al., 2018†	Voxel-GAN		Yes		Yes				Gaussian
Sun et al., 2018	CNN, DFKZ, 3D CNN	Yes							Gaussian
Wang et al., 2018*†	CNN	Yes	Yes		Yes				Random
Number of methods utilizing this augmentation →		15	8	2	9	1	2	1	8
Percentage (%) of methods utilizing this augmentation →		75	40	10	45	5	10	5	40

The top-performing techniques (over the unseen test set) are annotated with green.

^*The authors verified the impact of data augmentation of the generalization abilities of their deep models.

^†The authors used both training- and test-time data augmentation.