ninatubau / plant-seg

A tool for cell instance aware segmentation in densely packed 3D volumetric images

PlantSeg

PlantSeg is a tool for cell instance aware segmentation in densely packed 3D volumetric images. The pipeline uses a two stages segmentation strategy (Neural Network + Segmentation). The pipeline is tuned for plant cell tissue acquired with confocal and light sheet microscopy. Pre-trained models are provided.

Getting Started

The recommended way of installing plantseg is via the conda package, which is only supported on Linux. Running plantseg on other operating systems (Windows 10, Mac OS) is currently possible only via a Docker image that we provide (see below).

Prerequisites for conda package

• Linux
• (Optional) Nvidia GPU + CUDA

or

• Windows 10 / MacOS via Docker

Install Anaconda python

• First step required to use the pipeline is installing anaconda python. If you already have a working anaconda setup you can go directly to next item. To download Anaconda Python open a terminal and type

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

Then install by typing:

bash ./Miniconda3-latest-Linux-x86_64.sh

Follow the instructions to complete the anaconda installation.

Install PlantSeg using conda

The tool can be installed directly by executing in the terminal:

conda create -n plant-seg -c pytorch -c conda-forge -c lcerrone -c abailoni -c cpape -c awolny pytorch nifty=1.0.9 plantseg

Above command will create new conda environment plant-seg together with all required dependencies.

Optional dependencies

Some types of compressed tiff files require an additional package to be read correctly (eg: Zlib, ZSTD, LZMA, ...). To run plantseg on those stacks you need to install imagecodecs. In the terminal:

conda activate plant-seg
pip install imagecodecs

Experimental support for SimpleITK watershed segmentation has been added to PlantSeg version 1.1.8. This features can be used only after installing the SimpleITK package:

conda activate plant-seg
pip install SimpleITK

Update PlantSeg

The tool is actively developed and improvement and small bugs fix are continuosly coming. To update PlantSeg type in the terminal:

conda activate plant-seg
conda update -c lcerrone plantseg

Pipeline Usage (command line)

Our pipeline is completely configuration file based and does not require any coding.

First, activate the newly created conda environment with:

conda activate plant-seg

then, one can just start the pipeline with

plantseg --config CONFIG_PATH

where CONFIG_PATH is the path to the YAML configuration file. See config.yaml for a sample configuration file and detailed description of the parameters.

Pipeline Usage (GUI)

PlantSeg app can also be started in a GUI mode, where basic user interface allows to configure and run the pipeline. First, activate the newly created conda environment with:

conda activate plant-seg

then, run the GUI by simply typing:

$ plantseg --gui

Guide to Custom Configuration File

The configuration file defines all the operations in our pipeline together with the data to be processed. Please refer to config.yaml for a sample configuration of the pipeline and detailed explanation of all of the parameters.

Some key design choices:

• path attribute: is used to define either the file to process or the directory containing the data.
• preprocessing attribute: contains a simple set of possible operations one would need to run on their own data before calling the neural network. If data is ready for neural network processing either delete the entire section or set state: False in order to skip this step. Detailed instructions can be found at Data Processing.
• cnn_prediction attribute: contains all parameters relevant for predicting with neural network. Description of all pre-trained models provided with the package are described below. Detailed instructions can be found at Predictions.
• segmentation attribute: contains all parameters needed to run the partitioning algorithm (i.e. final segmentation). Detailed instructions can be found at Segmentation.

Additional information

The PlantSeg related files (models, configs) will be placed inside your home directory under ~/.plantseg_models.

Our pipeline uses the PyTorch library for the CNN predictions. PlantSeg can be run on systems without GPU, however for maximum performance we recommend that the application is run on a machine with a high performance GPU for deep learning. If CUDA_VISIBLE_DEVICES environment variable is not specified the prediction task will be distributed on all available GPUs. E.g. run: CUDA_VISIBLE_DEVICES=0 plantseg --config CONFIG_PATH to restrict prediction to a given GPU.

Repository index

The PlantSeg repository is organised as follows:

• plantseg: Contains the source code of PlantSeg.
• conda-reicpe: Contains all necessary code and configuration to create the anaconda package.
• Documentation-GUI: Contains an more in-depth documentation of PlantSeg functionality.
• evaluation: Contains all script required to reproduce the quantitative evaluation in Wolny et al..
• examples: Contains the files required to test PlantSeg.
• tests: Contains automated tests that ensures the PlantSeg functionality are not compromised during an update.

Docker image

We also provide a Docker image with plantseg package, which can be run on any operating system with Docker installed. Since plantseg is normally used in a GUI mode, one has to share a display on the host operating system with a docker container running plantseg. Below we provide a detailed instruction of how to run a plantseg Docker image on Linux, Windows 10 and Mac OS.

As a side note: running plantseg via Docker on Windows and Mac OS works only with CPU mode, which is significantly slower than when running on the GPU, e.g. for a 3D stack of size 200x400x400 it took ~40mins to segment with plantseg Docker image on Windows (as compared to 1.5 mins when segmenting the same stack using plantseg with GPU) on a modern laptop.

Also bear in mind that plantseg is quite memory hungry, so when running with Docker on a laptop, please make sure to process smaller volumes (up to 1GB) and use smaller patch sizes for neural network predictions, otherwise your Docker container may be terminated abruptly due to the out of memory issue.

Linux

Make sure that nvidia-docker is installed on the docker host otherwise you won't be able to utilize the GPUs.

In oder to execute the docker image in the GUI mode, fist we need to allow everyone to access X server on the docker host. This can be done by invoking the following command in the terminal:

xhost +

For GPU support run:

docker run --runtime=nvidia -it --rm -v PATH_TO_DATASET:/root/datasets -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=unix$DISPLAY wolny/plantseg

If your docker host does not have modern GPU and/or nvidia-docker is not installed, run:

docker run -it --rm -v PATH_TO_DATASET:/root/datasets -v /tmp/.X11-unix:/tmp/.X11-unix -e DISPLAY=unix$DISPLAY wolny/plantseg

this will start the plantseg GUI application. Replace PATH_TO_DATASET by the directory on the host machine that you want to mount (i.e. this is the path where your data is stored on your computer). /root/datasets is the directory inside the Docker container and this is where you should navigate the PlantSeg GUI to find your files.

Windows 10

• Install Docker Desktop on Windows. Make sure to have the latest Windows 10 version installed, if not you might need to sign up for the Windows Insider Program in order to install the version required by Docker.
• Install Windows X Server.
• After you install, configure and run the VcXsrv Windows X Server, open the Windows PowerShell and run:

set-variable -name DISPLAY -value YOUR-IP:0.0

replace YOUR-IP with your actual host IP address (you can find it by running ipconfig in the PowerShell)

• Run plantseg via:

docker run -it --rm -v PATH_TO_DATASET:/root/datasets -e DISPLAY=$DISPLAY wolny/plantseg

Replace PATH_TO_DATASET by the directory on the host machine that you want to mount (i.e. this is the path where your data is stored on your computer). /root/datasets is the directory inside the Docker container and this is where you should navigate the PlantSeg GUI to find your files. E.g. if your data is stored in C:\Datasets\ConfocalImages run:

docker run -it --rm -v C:\Datasets\ConfocalImages:/root/datasets -e DISPLAY=$DISPLAY wolny/plantseg

Mac OS

• Install Docker Desktop on Mac
• Install X Window System on Mac
• after you install and run XQuartz 2.7.10 on your Mac according to the instructions above, run:

docker run -it --rm -v PATH_TO_DATASET:/root/datasets -e DISPLAY=$DISPLAY wolny/plantseg

Replace PATH_TO_DATASET by the directory on the host machine that you want to mount (i.e. this is the path where your data is stored on your computer). /root/datasets is the directory inside the Docker container and this is where you should navigate the PlantSeg GUI to find your files. E.g. if your data is stored in /Users/foobar/datasets/confocal run:

docker run -it --rm -v /Users/foobar/datasets/confocal:/root/datasets -e DISPLAY=$DISPLAY wolny/plantseg

Troubleshooting

If you see the error:

_tkinter.TclError: couldn't connect to display "/private/tmp/com.apple.launchd.mAtyr7ic37/org.macosforge.xquartz:0

Please follow the instructions from kreshuklab#53 (comment)

Datasets

We publicly release the datasets used for training the networks which available as part of the PlantSeg package. Please refer to our publication for more details about the datasets:

• Arabidopsis thaliana ovules dataset (raw confocal images + ground truth labels)
• Arabidopsis thaliana lateral root (raw light sheet images + ground truth labels)

Both datasets can be downloaded from our OSF project

Pre-trained networks

The following pre-trained networks are provided with PlantSeg package out-of-the box and can be specified in the config file or chosen in the GUI.

• generic_confocal_3d_unet - alias for confocal_unet_bce_dice_ds2x see below
• generic_light_sheet_3d_unet - alias for lightsheet_unet_bce_dice_ds1x see below
• confocal_unet_bce_dice_ds1x - a variant of 3D U-Net trained on confocal images of Arabidopsis ovules on original resolution, voxel size: (0.235x0.075x0.075 µm^3) (ZYX) with BCEDiceLoss
• confocal_unet_bce_dice_ds2x - a variant of 3D U-Net trained on confocal images of Arabidopsis ovules on 1/2 resolution, voxel size: (0.235x0.150x0.150 µm^3) (ZYX) with BCEDiceLoss
• confocal_unet_bce_dice_ds3x - a variant of 3D U-Net trained on confocal images of Arabidopsis ovules on 1/3 resolution, voxel size: (0.235x0.225x0.225 µm^3) (ZYX) with BCEDiceLoss
• lightsheet_unet_bce_dice_ds1x - a variant of 3D U-Net trained on light-sheet images of Arabidopsis lateral root on original resolution, voxel size: (0.25x0.1625x0.1625 µm^3) (ZYX) with BCEDiceLoss
• lightsheet_unet_bce_dice_ds2x - a variant of 3D U-Net trained on light-sheet images of Arabidopsis lateral root on 1/2 resolution, voxel size: (0.25x0.325x0.325 µm^3) (ZYX) with BCEDiceLoss
• lightsheet_unet_bce_dice_ds3x - a variant of 3D U-Net trained on light-sheet images of Arabidopsis lateral root on 1/3 resolution, voxel size: (0.25x0.4875x0.4875 µm^3) (ZYX) with BCEDiceLoss
• confocal_2D_unet_bce_dice_ds1x - a variant of 2D U-Net trained on confocal images of Arabidopsis ovules. Training the 2D U-Net is done on the Z-slices (pixel size: 0.075x0.075 µm^3) with BCEDiceLoss
• confocal_2D_unet_bce_dice_ds2x - a variant of 2D U-Net trained on confocal images of Arabidopsis ovules. Training the 2D U-Net is done on the Z-slices (1/2 resolution, pixel size: 0.150x0.150 µm^3) with BCEDiceLoss
• confocal_2D_unet_bce_dice_ds3x - a variant of 2D U-Net trained on confocal images of Arabidopsis ovules. Training the 2D U-Net is done on the Z-slices (1/3 resolution, pixel size: 0.225x0.225 µm^3) with BCEDiceLoss
• lightsheet_unet_bce_dice_nuclei_ds1x - a variant of 3D U-Net trained on light-sheet images Arabidopsis lateral root nuclei on original resolution, voxel size: (0.235x0.075x0.075 µm^3) (ZYX) with BCEDiceLoss
• confocal_unet_bce_dice_nuclei_stain_ds1x - a variant of 3D U-Net trained on confocal images of Arabidopsis ovules nuclei stain on original resolution, voxel size: (0.35x0.1x0.1 µm^3) (ZYX) with BCEDiceLoss
• confocal_PNAS_3d - a variant of 3D U-Net trained on confocal images of shoot apical meristem dataset from: Jonsson, H., Willis, L., & Refahi, Y. (2017). Research data supporting Cell size and growth regulation in the Arabidopsis thaliana apical stem cell niche. https://doi.org/10.17863/CAM.7793. voxel size: (0.25x0.25x0.25 µm^3) (ZYX)
• confocal_PNAS_2d - a variant of 2D U-Net trained on confocal images of shoot apical meristem dataset from: Jonsson, H., Willis, L., & Refahi, Y. (2017). Research data supporting Cell size and growth regulation in the Arabidopsis thaliana apical stem cell niche. https://doi.org/10.17863/CAM.7793. pixel size: (25x0.25 µm^3) (YX)

Selecting a given network name (either in the config file or GUI) will download the network into the ~/.plantseg_models directory. Detailed description of network training can be found in our paper.

Training on New Data

For training new models we rely on the pytorch-3dunet. A similar configuration file can be used for training on new data and all the instructions can be found in the repo. When the network is trained it is enough to create ~/.plantseg_models/MY_MODEL_NAME directory and copy the following files into it:

• configuration file used for training: config_train.yml
• snapshot of the best model across training: best_checkpoint.pytorch
• snapshot of the last model saved during training: last_checkpoint.pytorch

The later two files are automatically generated during training and contain all neural networks parameters.

Now you can simply use your model for prediction by setting the model_name key to MY_MODEL_NAME.

If you want your model to be part of the open-source model zoo provided with this package, please contact us.

Using LiftedMulticut segmentation

As reported in our paper, if one has a nuclei signal imaged together with the boundary signal, we could leverage the fact that one cell contains only one nuclei and use the LiftedMultict segmentation strategy and obtain improved segmentation. We're going to use the Arabidopsis thaliana lateral root as an example. The LiftedMulticut strategy consist of running PlantSeg two times:

1. Using PlantSeg to predict the nuclei probability maps using the lightsheet_unet_bce_dice_nuclei_ds1x network. In this case only the pre-processing and CNN prediction steps are enabled in the config, see example config.

plantseg --config nuclei_predictions_example.yaml 

2. Using PlantSeg to segment the input image with LiftedMulticut algorithm given the nuclei probability maps from the 1st step. See example config. The notable difference is that in the segmentation part of the config we set name: LiftedMulticut and the nuclei_predictions_path as the path to the directory where the nuclei pmaps were saved in step 1. Also make sure that the path attribute points to the raw files containing the cell boundary staining (NOT THE NUCLEI).

plantseg --config lifted_multicut_example.yaml

If case when the nuclei segmentation is given, one should skip step 1., add is_segmentation=True flag in the config and directly run step 2.

Troubleshooting

• If you stumble in the following error message:

AssertionError:
The NVIDIA driver on your system is too old (found version xxxx).
Please update your GPU driver by downloading and installing a new
version from the URL: http://www.nvidia.com/Download/index.aspx
Alternatively, go to: http://pytorch.org to install
a PyTorch version that has been compiled with your version
of the CUDA driver.

or:

    raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

It means that your cuda installation does not match the default in plantseg. You can check your current cuda version by typing in the terminal

cat /usr/local/cuda/version.txt

Then you can re-install the pytorch version compatible with your cuda by activating your plant-seg environment:

conda activate plant-seg

and

conda install -c pytorch torchvision cudatoolkit=<YOU_CUDA_VERSION> pytorch

e.g. for cuda 9.2

conda install -c pytorch torchvision cudatoolkit=9.2 pytorch

Alternatively one can create the plant-seg environment from scratch and ensuring the correct version of cuda/pytorch, by:

conda create -n plant-seg -c lcerrone -c abailoni -c cpape -c awolny -c conda-forge cudatoolkit=<YOU_CUDA_VERSION> plantseg

• If you use plantseg from the GUI and you receive an error similar to:

RuntimeError: key : 'crop_volume' is missing, plant-seg requires 'crop_volume' to run

(or a similar message for any of the other keys) It might be that the last session configuration file got corrupted or is outdated. You should be able to solve it by removing the corrupted file config_gui_last.yaml.

If you have a standard installation of plantseg, you can remove it by executing on the terminal:

$ rm ~/.plantseg_models/configs/config_gui_last.yaml

• If you use plantseg from the comand line and you receive an error similar to:

RuntimeError: key : 'crop_volume' is missing, plant-seg requires 'crop_volume' to run

Please make sure that your configuratiuon has the correct formatting and contains all required keys. An updated example can be found inside the directory examples, in this repository.

• If when trying to execute the Lifted Multicut pipeline you receive an error like:

'cannot import name 'lifted_problem_from_probabilities' from 'elf.segmentation.features''

The solution is to re-install elf via

conda install -c conda-forge -c cpape elf

• PlantSeg is under active development so it may happen that the models/configuration files saved in ~/.plantseg_modes are outdated. In case of errors related to loading the configuration file, please close the PlantSeg app, remove ~/.plantseg_models directory and try again.

Tests

In order to run tests make sure that pytest is installed in your conda environment. You can run your tests simply with python -m pytest or pytest. For the latter to work you need to install plantseg locally in "develop mode" with pip install -e ..

Citation

@article{wolny2020accurate,
  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, S{\"o}ren and Wilson-S{\'a}nchez, David and Lymbouridou, Rena and others},
  journal={Elife},
  volume={9},
  pages={e57613},
  year={2020},
  publisher={eLife Sciences Publications Limited}
}