TABLE III.

Comparisons of the Eight Top-Ranked Methods in the iSeg-2019 Challenge

TEAM	Architecture	Tool	How to deal with site differences?	Key highlight	Augmentation	Training Loss	2D/3D Patch size (training/testing)
QL111111	Self-attention and multi-scale dilated 3D UNet	Tensorflow	Intensity adjustment	Attention mechanism +multi-scale	No	Cross-entropy	3D (32×32×32)
Tao_SMU	Attention-guided Full-resolution Network	Tensorflow	Training data augmentation, contrast and lightness adjustment	Full resolution + attention mechanism	Dimension transformation, contrast and lightness adjusting	Cross-entropy	3D (32×32×32)
FightAutism	3D UNet	Pytorch	Histogram matching	Automate designing of segmentation pipeline	Random rotation, scaling, mirroring, and gamma transformation.	Cross-entropy	3D (112×128×128)
xflz	Intensity-augmented 3D UNet	Tensorflow	Intensity adjustment	Intensity augmentation for adaptation of multi-site data	Intensity augmentation, gaussian nosie and flip	Cross-entropy	3D (64×64×64)
SmartDSP	Adversarial learning 3D UNet	Pytorch	Global feature alignment	Adversarial learning + automate designing of segmentation pipeline	-	Cross-entropy, dice loss and adversarial loss	3D (112×128×128)
CU_SIAT	Entropy Minimization 3D densely connected network	Pytorch	Distribution alignment	Adversarial entropy minimization strategy	-	Cross-entropy and adversarial loss	3D (64×64×64)
trung	Cross-linkedFC-DenseNet	Pytorch	-	Cross link + channel attention	-	Cross-entropy	3D (64×64×64)
RB	Dense residual 3D Unet	Pytorch	-	Dense block + residual connection	-	Cross-entropy and dice loss	3D (64×64×64)