dlab-arp / cellseg-transformers

Multi-Modality Cell Segmentation

CellSeg-Transformers

Scripts to reproduce the results in the response to "Transformers do not outperform Cellpose"

Installation

# create virtual environment
conda create --name cp3 python=3.10 -y
conda activate cp3

# install the latest cellpose
git clone https://github.com/mouseland/cellpose.git
cd cellpose
pip install -e .

# install GPU version of torch
pip uninstall torch
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# install segmentation_models_pytorch for Transformers
pip install segmentation_models_pytorch
pip install six pandas

Fig1c: comparison between w/TTA and w/o TTA

python infer_cp_noTTA.py

Fig2a: Transformer model with learning rate 0.00005

1. Generate the flows: python gen_flow
2. Run the training command: cp-trans224-5e-5.sh
3. Download the pre-trained model here and run the inference script: python infer_new_trans_neurips_data.py
4. Submit the results to the challenge platform

Fig2b-m: New experiments on CTC cell segmentation dataset

0. Download the organized dataset here

1. Infer CTC dataset with Cellpose and Cellpose-Transformerd trained by Dr. Carsen Stringer and Dr. Marius Pachitariu: python infer_ctc492.py

2. Infer CTC dataset with Mediar, which was the winning solution in the NeurIPS 2022 segmentation challenge

   • Download the docker here
   • Run the inference docker container run --gpus="device=0" -m 28G --name mediar --rm -v $PWD/CTC-Data/imagesTr_GT492/:/workspace/inputs/ -v $PWD/CTC-Data/seg_mediar:/workspace/outputs/ osilab:latest /bin/bash -c "sh predict.sh"

3. Compute Metrics: python compute_metrics -g path_to_gt -s path_to_seg -o save_path -n save_name

Running time comparison for whole-slide images (WSI) segmenation

• Infer WSIs with Cellpose (Video)

python infer_wsi_cellpose_time.py

• Infer WSIs with Mediar (Video)

python infer_wsi_mediar_time.py