• Introduction
• Installation
• Getting Started
• Citation
• Issues

This repository hosts the version of the code used for the preprint titled Unsupervised Learning of Object-Centric Embeddings for Cell Instance Segmentation in Microscopy Images. This work was accepted to the International Conference for Computer Vision (ICCV), 2023.

We refer to the proposed techniques described in the preprint as Cellulus - Cellulus is a deep learning based method which can be used to obtain instance-segmentation of objects in 2D or 3D microscopy images in an unsupervised fashion i.e. requiring no ground truth labels during training.

One could execute these lines of code below to create a new environment and install dependencies.

1. Create a new environment called cellulus:

conda create -y -n cellulus python==3.9

2. Activate the newly-created environment:

conda activate cellulus

3a. If using a GPU, install pytorch cuda dependencies:

conda install pytorch==2.0.1 torchvision==0.15.2 pytorch-cuda=11.7 -c pytorch -c nvidia

3b. otherwise (if using a CPU or MPS), run:

pip install torch torchvision

4. Install the package from github:

pip install git+https://github.com/funkelab/cellulus.git

Try out a 2D example or a 3D example available under the examples tab here.

Using cellulus from the terminal window requires specifying a train.toml config file and an infer.toml config file.
These files indicate how the training and inference should be performed respectively.

For example, a minimal train.toml config file would look as follows:

[model_config]

num_fmaps = 256
fmap_inc_factor = 3
downsampling_factors = [[2,2],]

[train_config.train_data_config]

container_path = "skin.zarr" # specify path to zarr container, containing raw image dataset
dataset_name = "train/raw"

The train.toml recipe file can then be used to initiate the model training by running the following line in the terminal window:

train train.toml

Similarly, a minimal infer.toml file would look as follows:

[model_config]

num_fmaps = 256
fmap_inc_factor = 3
checkpoint = "models/best_loss.pth" # path to model weights

[inference_config.dataset_config]

container_path = "skin.zarr" # specify path to zarr container, containing raw image dataset
dataset_name = "test/raw"

[inference_config.prediction_dataset_config]

container_path = "skin.zarr"
dataset_name = "embeddings"

[inference_config.detection_dataset_config]

container_path = "skin.zarr"
dataset_name = "detection"
secondary_dataset_name = "embeddings"

[inference_config.segmentation_dataset_config]

container_path = "skin.zarr"
dataset_name = "segmentation"
secondary_dataset_name = "detection"

The infer.toml recipe file can be used to apply the trained model weights on raw image data and obtain instance segmentations, by running the following line in the terminal window:

infer infer.toml

If you find our work useful in your research, please consider citing:

@misc{wolf2023unsupervised,
      title={Unsupervised Learning of Object-Centric Embeddings for Cell Instance Segmentation in Microscopy Images},
      author={Steffen Wolf and Manan Lalit and Henry Westmacott and Katie McDole and Jan Funke},
      year={2023},
      eprint={2310.08501},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

If you encounter any problems, please file an issue along with a description.