A Large-Scale Multimodal Annotated Dataset for Cell Segmentation with Benchmarking of Universal Models

Introduction

This project is a benchmark for general cell segmentation models. We have deployed the following 7 cell segmentation methods: MEDIAR, Cellpose, Cellpose3, SAM, Stardist, Deepcell, Cellprofiler, along with code to evaluate the model segmentation performance. Through the command line or notebook, you can run 7 segmentation models or evaluate the performance of 7 models with one click.
Here is an evaluation example of DAPI staining in the DEMO:

Installation

git clone https://github.com/STOmics/cs-benchmark.git   
cd cs-benchmark 

• Create an environment for Cellpose, SAM, StarDist, and DeepCell.
• NOTE：The command does not include installing PyTorch. If you need to use GPU, please install the corresponding version of PyTorch.

# python3.8 in conda env
conda create --name=cs-benchmark python=3.8
conda activate cs-benchmark
pip install -r requirements.txt
pip install git+https://github.com/facebookresearch/segment-anything.git

• Use the following command to install the environment for mediar and cellprofiler, and add the conda path in the _py_ section of cell_seg.py or cellsegmentation_benchmark.ipynb

conda env create -f src/methods/MEDIAR/environment.yaml
conda env create -f src/methods/cellprofiler/environment.yaml

• Download the necessary model file

  sam_vit_b_01ec64.pth
  MEDIAR
  After downloading the model file, place it in the src/methods/models directory of the project.

Tutorials

Data

zenodo download link
CNSA download link

Use via command line

Cell segmentation

• Modify the parameters in the following command and input it into the command line:

python cell_seg.py -i your_inputpath -o your_outputpath -m  cellpose3 sam -t ss -g True  

• Where:

• -i is the input image path
  -o is the output mask path
  -m is the algorithm(s) to be used (can specify multiple)
  -t is the image type (ss/he/dapi/mif)
  -g is the GPU index (True/False or num)

Segmentation evaluation

• Ensure that the images in the gt folder have filenames with "-mask" and the images in the algorithm output mask folder have filenames with "-img", with only this difference in their names.

• Modify the parameters in the following command and input it into the command line:

python src/eval/cell_eval_multi.py -g gt_path -d dt_path -o result_path

• Where:

• -g is the path to the ground truth (GT) folder
  -d is the path to the algorithm output mask folder
  -o is the output path for the results

Use via Notebook

cellsegmentation_benchmark.ipynb

Citation

if you use CellBinDB in your work, please cite it

Shi C, Fan J, Deng Z, et al. CellBinDB: A Large-Scale Multimodal Annotated Dataset for Cell Segmentation with Benchmarking of Universal Models[J]. bioRxiv, 2024: 2024.11. 20.619750.
doi: https://doi.org/10.1101/2024.11.20.619750

Reference

cellpose
cellpose3
deepcell
sam
mediar
stardist
cellprofiler