NeurIPS22-CellSeg 0.0.1

NeurIPS-CellSeg-baseline

NeurIPS-CellSeg

A naive baseline and submission demo for the microscopy image segmentation challenge in NeurIPS 2022

Requirements

MONAI version: 0.9
Numpy version: 1.21.2
Pytorch version: 1.10.1
Nibabel version: 3.2.1
scikit-image version: 0.19.2
Pillow version: 9.0.1
Tensorboard version: 2.8.0
gdown version: 4.2.0
TorchVision version: 0.11.2
tqdm version: 4.63.0
psutil version: 5.8.0
pandas version: 1.4.1
einops version: 0.3.2

Preprocessing

Download training data to the data folder

Run python pre_process_3class.py

Training

cd baseline

Run python model_training_3class.py --data_path 'path to training data' --batch_size 8

Inference

Run

python predict.py -i input_path -o output_path

Your prediction file should have at least the two arguments: input_path and output_path. The two arguments are important to establish connections between local folders and docker folders.

Build Docker

We recommend this great tutorial: https://nbviewer.org/github/ericspod/ContainersForCollaboration/blob/master/ContainersForCollaboration.ipynb

1) Preparation

The docker is built on MONAI

docker pull projectmonai/monai

Prepare Dockerfile

FROM projectmonai/monai:latest

WORKDIR /workspace
COPY ./   /workspace

Put the inference command in the predict.sh

# !/bin/bash -e
python predict.py -i "/workspace/inputs/"  -o "/workspace/outputs/"

The input_path and output_path augments should specify the corresponding docker workspace folders rather than local folders, because we will map the local folders to the docker workspace folders when running the docker container.

2) Build Docker and make sanity test

The submitted docker will be evaluated by the following command:

docker container run --gpus "device=0" --name teamname --rm -v $PWD/CellSeg_Test/:/workspace/inputs/ -v $PWD/teamname_outputs/:/workspace/outputs/ teamname:latest /bin/bash -c "sh predict.sh"

• --name: container name during running

• --rm: remove the container after running

• -v $PWD/CellSeg_Test/:/workspace/inputs/: map local image data folder to Docker workspace/inputs folder.

• -v $PWD/teamname_outputs/:/workspace/outputs/ : map Docker workspace/outputs folder to local folder. The segmentation results will be in $PWD/teamname_outputs

• teamname:latest: docker image name (should be teamname) and its version tag. The version tag should be latest. Please do not use v0, v1... as the version tag

• /bin/bash -c "sh predict.sh": start the prediction command. It will load testing images from workspace/inputs and save the segmentation results to workspace/outputs

Assuming the team name is baseline, the Docker build command is

docker build -t baseline .

Test the docker to make sure it works. There should be segmentation results in the baseline_outputs folder.

docker container run --gpus "device=0" --name baseline --rm -v $PWD/CellSeg_Test/:/workspace/inputs/ -v $PWD/baseline_outputs/:/workspace/outputs/ baseline:latest /bin/bash -c "sh predict.sh"

During the inference, please monitor the GPU memory consumption using watch nvidia-smi. The GPU memory consumption should be less than 1500MB. Otherwise, it will run into an OOM error on the official evaluation server. We impose this hard constraint on GPU memory consumption to ensure ease of use, because biologists may not have powerful GPUs in practice. Thus, the model should be low-resource.

3) Save Docker

docker save baseline | gzip -c > baseline.tar.gz

Upload the docker to Google drive or Baidu net disk and send the download link to [e-mail].

Please do not upload the Docker to dockerhub!

Limitations and potential improvements

The naive baseline's primary aim is to give participants out-of-the-box scripts that can generate successful submisions. Thus, there are many ways to surpass this baseline:

• New cell representation methods. In the baseline, we separated touching cells by simply removing their boundaries. More advanced cell representation could be used to address this issue, such as stardist, cellpose, omnipose, deepcell, and so on.
• New architectures
• More data augmentations and the use of additional public datasets or the set of unlabeled data provided.
• Well-designed training protocols
• Postprocessing

Nevertheless, please always keep in mind that many end users do not have powerful computation resources. It's important to consider the trade-off between resource consumption and accuracy.