Unsupervised Nuclei Segmentation using Spatial Organization Priors

Code for training and benchmarking unsupervised and supervised methods on several dataset of H-DAB stained immunohistochemistry images for nuclei segmentation.

• train_GAN.py trains the proposed cycle GAN model for a given IHC dataset and ground truth masks of nuclei segmentation from any source.
• train_Unet.py trains the generator of the proposed method (based on an Unet architecture) in a supervised fashion
• tuning_Unet.py retrieves best set of hyperparameters for training the generator in a supervised fashion.
• test.py computes every metrics for all the benchmarked methods.

Installation

The packages needed to run the code are provided in environment.yml. The environment can be recreated using:

conda env create -f environment.yml

Data

4 datasets were used in this paper:

• Deepliif: Set of 1667 registered 512x512 IHC images built to train the DEEPLIIF algorithm; only the original Ki67 stained and the nuclei segmentation images were used - see here for the data.
• Warwick: Set of 86 HER2 stained WSIs - see here for more details and request access to the data.
• BCDataset: Set of 1338 640x640 IHC images stained for Ki67 with cell center annotations - see here for more details and retrieve the data.
• Pannuke: Set of more than 7000 256x256 H&E stained images with corresponding nuclei segmentation - see here for more details and retrieve the data.

Please see the directory dataset for more information about how was used each dataset.

Proposed method

train_GAN.py requires 3 arguments to train the GAN model:

• dataihcroot_images describes the location of H-DAB stained training data.
• dataroot_masks describes the location of nuclei segmentation masks from another database (Pannuke here).
• dataset_name specifies which dataset is used among Deepliif and Warwick (only these 2 were used in this paper for training).

and some optional arguments are available to specify the number of epochs, batch_size, number of worker nodes, ...

Example command line:

python train.py /path/to/DeepLiif/ /path/to/PanNuke/ deepliif

test.py requires 4 arguments to compute performance metrics:

• data_path describes the location of H-DAB stained testing data (H-DAB stained images and their corresponding nuclei segmentation)
• method specifies the method to benchmark ("GAN", "Unet", "Stardist", "Nuclick", "Thresholding")
• dataset_name specifies which dataset is used among Deepliif and Warwick (only these 2 were used in this paper to compute metrics).
• Depending on the chosen method, --benchmark_path or --checkpoint_path must be provided:
  • --checkpoint_path describes the location of best saved models for "GAN" and "Unet" methods. It must contain model checkpoints and training losses for "GAN" to select the best model over training iterations. If already selected, you may just indicate the path to the model, named: "proposed_{deepliif/warwick}.pt" as it is done for provided pretrained models. It must contain optimal hyperparameters for "Unet".
  • --benchmark_path describes the location of the mask computed with existing methods ("Stardist", "Nuclick", "Thresholding")

Example command line:

python test.py /path/to/DeepLiif/ GAN deepliif --checkpoint_path /path/to/training_results --batch_size=46

python test.py /path/to/Warwick/ Nuclick warwick --benchmark_path /path/to/Nuclick_results

The pretrained models can be retrieved here.

Existing methods

3 methods were used to benchmark the proposed method:

• Qupath software with Stardist plugin. The scripts are available here.
• Nuclick algorithm to extract mask segmentation from nuclei center annotations. The code of Nuclick can be retrieved here, and the code for this paper here.
• A baseline thresholding approach, using color deconvolution for membranous staining. The code is available here

Acknowledgement

Many thanks to Zhu.et.al for providing the code of their cycleGAN architectures - see here.

Reference

If you find this code useful in your research then please cite:

@inproceedings{lebescond:hal-03644463,
  TITLE = {{Unsupervised Nuclei Segmentation using Spatial Organization Priors}},
  AUTHOR = {Le Bescond, Lo{\"i}c and Lerousseau, Marvin and Garberis, Ingrid and Andr{\'e}, Fabrice and Christodoulidis, Stergios and Vakalopoulou, Maria and Talbot, Hugues},
  BOOKTITLE = {{MICCAI 2022 - 25th International Conference on Medical Image Computing and Computer Assisted Intervention}},
  YEAR = {2022},
}

License

This code is MIT licensed, as found in the LICENSE file.