Skip to content

Latest commit

 

History

History
 
 

test

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
data_splits
data_splits
 
 
nnUNetV1
nnUNetV1
 
 
nnUNetV2
nnUNetV2
 
 
LossOverview.PNG
LossOverview.PNG
 
 
README.md
README.md
 
 

README.md

Usage Instructions

All the loss functions have been tested with the nnUNetTrainer in nnUNet V1.

  1. Prerequisites: install nnUNet.
  2. Download the loss functions: git clone https://github.com/JunMa11/SegLoss.git
  3. Copy SegLoss/test/nnUNetV1/loss_functions and SegLoss/test/nnUNetV1/network_training to nnUNet/nnunet/training
  4. To Train your model, replacing nnUNetTrainer by the new trainer. e.g., if you want to train UNet with Dice loss, run:

python run/run_training.py 3d_fullres nnUNetTrainer_Dice TaskXX_MY_DATASET FOLD --ndet

Datasets

Many loss functions can achieve great performance on liver segmentation. Thus, we recommend paying more attention to the other three tasks.

Results

Download associated segmentation results

Loss Liver-DSC Liver-NSD Liver Tumor-DSC Liver Tumor-NSD Pancreas-DSC Pancreas-NSD Multi-organ-DSC Multi-organ-NSD
Asym 0.9315 0.6905 0.6134 0.4114 0.8234 0.6239 0.7526 0.6088
CE 0.9672 0.7962 0.5641 0.3715 0.8221 0.6321 0.8483 0.7200
Dice 0.9547 0.7598 0.6187 0.4291 0.8362 0.6688 0.8449 0.7136
DiceBD 0.9629 0.7702 0.6262 0.4375 0.8397 0.6713 0.8450 0.7105
DiceCE 0.9624 0.7743 0.6009 0.4114 0.8249 0.6298 0.8512 0.7293
DiceFocal 0.9566 0.7583 0.5951 0.4078 0.8416 0.6721 0.8554 0.7339
DiceHD 0.9556 0.7314 0.6291 0.4390 0.8408 0.6646 0.8531 0.7257
DiceTopK 0.9690 0.8092 0.6125 0.4208 0.8375 0.6598 0.8512 0.7308
ELL 0.9347 0.7254 0.5903 0.4047 0.8344 0.6508 0.8375 0.6689
Focal 0.9587 0.7528 0.4675 0.2781 0.8016 0.6034 0.8173 0.6642
FocalTversky 0.9320 0.6986 0.6193 0.4178 0.8229 0.6190 0.7497 0.6013
GDice 0.9474 0.7337 0.5486 0.3501 0.8285 0.6478 0.0132 0.0018
IoU 0.9568 0.7709 0.6079 0.4273 0.8353 0.6605 0.8439 0.7160
Lovasz 0.9294 0.6639 0.6083 0.4205 0.8309 0.6521 0.6568 0.3845
pCE 0.9655 0.7852 0.5876 0.3829 0.8358 0.6580 0.8349 0.6967
pGDice 0.6449 0.3455 0.4526 0.2879 0.8156 0.6204 0.0595 0.0209
SS 0.9591 0.7571 0.4527 0.1941 0.7799 0.4781 0.7589 0.4958
TopK 0.6924 0.1073 0.5995 0.4051 0.8406 0.6709 0.8527 0.7323
Tversky 0.9390 0.6991 0.6120 0.4045 0.8260 0.6249 0.8371 0.6787
WCE 0.8284 0.2665 0.2697 0.0314 0.4744 0.0496 0.6904 0.2335

To Do

  • Evaluate commonly used plug-and-play loss functions with nnU-Net V2 on three label-imbalanced tasks (liver tumor, pancreas, multi-organ) because the latest version is more popular (and has better performance).

In nnU-Net V2, deep supervision is added to the default U-Net. The optimizer is SGD with momentum rather than Adam.

The associated segmentation results have been released.

Loss LiverTumor-DSC LiverTumor-NSD Pancreas-DSC Pancreas-NSD Multiorgan-DSC Multiorgan-NSD
CE 0.6415 0.4698 0.8338 0.6566 0.8570 0.7368
Dice 0.6200 0.4592 0.8399 0.6663 0.8577 0.7416
DiceCE 0.6281 0.4678 0.8410 0.6691 0.8626 0.7488
DiceFocal 0.6303 0.4705 0.8401 0.6691 0.8631 0.7501
DiceTopK10 0.6691 0.5095 0.8387 0.6661 0.8636 0.7483
TopK10 0.6512 0.4849 0.8383 0.6649 0.8560 0.7378