Table 1.

The detailed configuration for multi-resolution segmentation framework

Procedure	Design choice	Coarse model	Fine model
Pre-processing	Intensity normalization	If CT, z score with fixed mean and standard deviation (SD) & clipping to [−1, 1]; If MRI, percentile z score with mean and SD & clipping to [−1, 1]	If CT, z score with fixed mean and SD & clipping to [−1, 1]; If MRI, percentile z score with mean and SD & clipping to [−1, 1]
	Image resampling strategy	Nearest neighbor interpolation	Nearest neighbor interpolation; Linear interpolation
	Annotation resampling strategy	[0, 1, …, class-1] encoding nearest neighbor / linear interpolation	[0, 1, …, class-1] encoding nearest neighbor interpolation
	Image target spacing	Spacing fixed to [5, 5, 5]	Spacing fixed to [1, 1, 1]
	Network topology	VB-Net for common organs; Adaptive VB-Net for large organs	VB-Net for common organs; Adaptive VB-Net for large organs
	Patch size	[96, 96, 96]	[96, 96, 96] for common organs; [196, 196, 196] for large organs
	Batch size	At least 2, given multi-GPU memory constraint	At least 2, given multi-GPU memory constraint
Training	Learning rate	Step learning rate schedule (initial, 1e-4)	Step learning rate schedule (initial, 1e-4)
	Loss function	Dice and boundary Dice	Dice and boundary Dice for OARs; 3D Dice, boundary Dice, and adaptive 2D Dice for CTV and PTV
	Optimizer	Adam (momentum = 0.9, decay = 1e-4, betas = (0.9, 0.999))	Adam (momentum = 0.9, decay = 1e-4, betas = (0.9, 0.999))
	Data augmentation	Rotating, scaling, flipping, shifting & adding noise	Rotating, scaling, flipping, shifting & adding noise
	Training procedure	1000 epochs, global sampling & mask sampling	1000 epochs, global sampling & mask sampling
Testing	Configuration for pre-processing	Resampling to fix spacing as training; Image partition given GPU memory	Available to expand the bounding box with user-set size or not; Resampling to fix spacing as training; Image partition given GPU memory
	Configuration for post-processing	Resampling to raw image spacing; Available to pick the largest connected component (CC) in segmentation or not; Available to remove small CC in segmentation or not	Resampling to raw image spacing; Available to pick the largest CC in segmentation or not; Available to remove small CC in segmentation or not