PyTorch implementation 3D U-Net and its variants:
-
Standard 3D U-Net based on 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation Özgün Çiçek et al.
-
Residual 3D U-Net based on Superhuman Accuracy on the SNEMI3D Connectomics Challenge Kisuk Lee et al.
The code allows for training the U-Net for both: semantic segmentation (binary and multi-class) and regression problems (e.g. de-noising, learning deconvolutions).
Training the standard 2D U-Net is also possible, see 2DUnet_dsb2018 for example configuration. Just make sure to keep the singleton z-dimension in your H5 dataset (i.e. (1, Y, X)
instead of (Y, X)
) , because data loading / data augmentation requires tensors of rank 3 always.
- Linux
- NVIDIA GPU
- CUDA CuDNN
The package has not been tested on Windows, however some users reported using it successfully on Windows.
- BCEWithLogitsLoss (binary cross-entropy)
- DiceLoss (standard
DiceLoss
defined as1 - DiceCoefficient
used for binary semantic segmentation; when more than 2 classes are present in the ground truth, it computes theDiceLoss
per channel and averages the values) - BCEDiceLoss (Linear combination of BCE and Dice losses, i.e.
alpha * BCE + beta * Dice
,alpha, beta
can be specified in theloss
section of the config) - CrossEntropyLoss (one can specify class weights via the
weight: [w_1, ..., w_k]
in theloss
section of the config) - PixelWiseCrossEntropyLoss (one can specify per pixel weights in order to give more gradient to the important/under-represented regions in the ground truth)
- WeightedCrossEntropyLoss (see 'Weighted cross-entropy (WCE)' in the below paper for a detailed explanation)
- GeneralizedDiceLoss (see 'Generalized Dice Loss (GDL)' in the below paper for a detailed explanation) Note: use this loss function only if the labels in the training dataset are very imbalanced e.g. one class having at least 3 orders of magnitude more voxels than the others. Otherwise use standard DiceLoss.
For a detailed explanation of some of the supported loss functions see: Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations Carole H. Sudre et al.
- MSELoss (mean squared error loss)
- L1Loss (mean absolute errro loss)
- SmoothL1Loss (less sensitive to outliers than MSELoss)
- WeightedSmoothL1Loss (extension of the SmoothL1Loss which allows to weight the voxel values above/below a given threshold differently)
- MeanIoU (mean intersection over union)
- DiceCoefficient (computes per channel Dice Coefficient and returns the average) If a 3D U-Net was trained to predict cell boundaries, one can use the following semantic instance segmentation metrics (the metrics below are computed by running connected components on thresholded boundary map and comparing the resulted instances to the ground truth instance segmentation):
- BoundaryAveragePrecision (Average Precision applied to the boundary probability maps: thresholds the output from the network, runs connected components to get the segmentation and computes AP between the resulting segmentation and the ground truth)
- AdaptedRandError (see http://brainiac2.mit.edu/SNEMI3D/evaluation for a detailed explanation)
- AveragePrecision (see https://www.kaggle.com/stkbailey/step-by-step-explanation-of-scoring-metric)
If not specified MeanIoU
will be used by default.
- PSNR (peak signal to noise ratio)
- MSE (mean squared error)
- The easiest way to install
pytorch-3dunet
package is via conda:
conda create -n pytorch3dunet -c pytorch -c conda-forge -c awolny pytorch-3dunet
conda activate pytorch3dunet
After installation the following commands are accessible within the conda environment:
train3dunet
for training the network and predict3dunet
for prediction (see below).
- One can also install directly from source:
python setup.py install
Make sure that the installed pytorch
is compatible with your CUDA version, otherwise the training/prediction will fail to run on GPU. You can re-install pytorch
compatible with your CUDA in the pytorch3dunet
environment by:
conda install -c pytorch cudatoolkit=<YOU_CUDA_VERSION> pytorch
Given that pytorch-3dunet
package was installed via conda as described above, one can train the network by simply invoking:
train3dunet --config <CONFIG>
where CONFIG
is the path to a YAML configuration file, which specifies all aspects of the training procedure.
In order to train on your own data just provide the paths to your HDF5 training and validation datasets in the config.
- sample config for 3D semantic segmentation (cell boundary segmentation): train_config_segmentation.yaml)
- sample config for 3D regression task (denoising): train_config_regression.yaml)
The HDF5 files should contain the raw/label data sets in the following axis order: DHW
(in case of 3D) CDHW
(in case of 4D).
One can monitor the training progress with Tensorboard tensorboard --logdir <checkpoint_dir>/logs/
(you need tensorflow
installed in your conda env), where checkpoint_dir
is the path to the checkpoint directory specified in the config.
- When training with binary-based losses, i.e.:
BCEWithLogitsLoss
,DiceLoss
,BCEDiceLoss
,GeneralizedDiceLoss
: The target data has to be 4D (one target binary mask per channel). When training withWeightedCrossEntropyLoss
,CrossEntropyLoss
,PixelWiseCrossEntropyLoss
the target dataset has to be 3D, see also pytorch documentation for CE loss: https://pytorch.org/docs/master/generated/torch.nn.CrossEntropyLoss.html final_sigmoid
in themodel
config section applies only to the inference time (validation, test): When training with cross entropy based losses (WeightedCrossEntropyLoss
,CrossEntropyLoss
,PixelWiseCrossEntropyLoss
) setfinal_sigmoid=False
so thatSoftmax
normalization is applied to the output. When training withBCEWithLogitsLoss
,DiceLoss
,BCEDiceLoss
,GeneralizedDiceLoss
setfinal_sigmoid=True
Given that pytorch-3dunet
package was installed via conda as described above, one can run the prediction via:
predict3dunet --config <CONFIG>
In order to predict on your own data, just provide the path to your model as well as paths to HDF5 test files (see example test_config_segmentation.yaml).
In order to avoid patch boundary artifacts in the output prediction masks the patch predictions are averaged, so make sure that patch/stride
params lead to overlapping blocks, e.g. patch: [64, 128, 128] stride: [32, 96, 96]
will give you a 'halo' of 32 voxels in each direction.
By default, if multiple GPUs are available training/prediction will be run on all the GPUs using DataParallel.
If training/prediction on all available GPUs is not desirable, restrict the number of GPUs using CUDA_VISIBLE_DEVICES
, e.g.
CUDA_VISIBLE_DEVICES=0,1 train3dunet --config <CONFIG>
or
CUDA_VISIBLE_DEVICES=0,1 predict3dunet --config <CONFIG>
Training/predictions configs can be found in 3DUnet_lightsheet_boundary. Pre-trained model weights available here. In order to use the pre-trained model on your own data:
- download the
best_checkpoint.pytorch
from the above link - add the path to the downloaded model and the path to your data in test_config.yml
- run
predict3dunet --config test_config.yml
- optionally fine-tune the pre-trained model with your own data, by setting the
pre_trained
attribute in the YAML config to point to thebest_checkpoint.pytorch
path
The data used for training can be downloaded from the following OSF project:
- training set: https://osf.io/9x3g2/
- validation set: https://osf.io/vs6gb/
- test set: https://osf.io/tn4xj/
Sample z-slice predictions on the test set (top: raw input , bottom: boundary predictions):
Training/predictions configs can be found in 3DUnet_confocal_boundary. Pre-trained model weights available here. In order to use the pre-trained model on your own data:
- download the
best_checkpoint.pytorch
from the above link - add the path to the downloaded model and the path to your data in test_config.yml
- run
predict3dunet --config test_config.yml
- optionally fine-tune the pre-trained model with your own data, by setting the
pre_trained
attribute in the YAML config to point to thebest_checkpoint.pytorch
path
The data used for training can be downloaded from the following OSF project:
- training set: https://osf.io/x9yns/
- validation set: https://osf.io/xp5uf/
- test set: https://osf.io/8jz7e/
Sample z-slice predictions on the test set (top: raw input , bottom: boundary predictions):
Training/predictions configs can be found in 3DUnet_lightsheet_nuclei. Pre-trained model weights available here. In order to use the pre-trained model on your own data:
- download the
best_checkpoint.pytorch
from the above link - add the path to the downloaded model and the path to your data in test_config.yml
- run
predict3dunet --config test_config.yml
- optionally fine-tune the pre-trained model with your own data, by setting the
pre_trained
attribute in the YAML config to point to thebest_checkpoint.pytorch
path
The training and validation sets can be downloaded from the following OSF project: https://osf.io/thxzn/
Sample z-slice predictions on the test set (top: raw input, bottom: nuclei predictions):
The data can be downloaded from: https://www.kaggle.com/c/data-science-bowl-2018/data
Training/predictions configs can be found in 2DUnet_dsb2018.
Sample predictions on the test image (top: raw input, bottom: nuclei predictions):
If you want to contribute back, please make a pull request.
If you use this code for your research, please cite as:
@article {10.7554/eLife.57613,
article_type = {journal},
title = {Accurate and versatile 3D segmentation of plant tissues at cellular resolution},
author = {Wolny, Adrian and Cerrone, Lorenzo and Vijayan, Athul and Tofanelli, Rachele and Barro, Amaya Vilches and Louveaux, Marion and Wenzl, Christian and Strauss, Sören and Wilson-Sánchez, David and Lymbouridou, Rena and Steigleder, Susanne S and Pape, Constantin and Bailoni, Alberto and Duran-Nebreda, Salva and Bassel, George W and Lohmann, Jan U and Tsiantis, Miltos and Hamprecht, Fred A and Schneitz, Kay and Maizel, Alexis and Kreshuk, Anna},
editor = {Hardtke, Christian S and Bergmann, Dominique C and Bergmann, Dominique C and Graeff, Moritz},
volume = 9,
year = 2020,
month = {jul},
pub_date = {2020-07-29},
pages = {e57613},
citation = {eLife 2020;9:e57613},
doi = {10.7554/eLife.57613},
url = {https://doi.org/10.7554/eLife.57613},
keywords = {instance segmentation, cell segmentation, deep learning, image analysis},
journal = {eLife},
issn = {2050-084X},
publisher = {eLife Sciences Publications, Ltd},
}