HASTE Team: Adipocyte Cell Challenge

In this repository we provide the code base of our solution to the challenge hosted by AI Sweden and Astra Zeneca (https://www.ai.se/en/news/challenge-can-machine-learning-replace-potential-cytotoxic-labeling-cell-cultures). The content and structure of the repo is given by the following:

.
├── README.md
├── ai_haste
│   ├── data                : python scripts for loading the datasets required
│   ├── loss                : for loss functions not included in PyTorch
│   ├── model               : python scripts for the various neural networks
│   ├── tester              : scripts requied when running the models in test mode
│   ├── trainer             : scripts requied when running the models in train mode
│   └── utils               : utility functions (e.g. for converting the images to numpy arrays for faster data loading)
├── config
│   ├── train               : .json files for running the models to reconstruct the three fluorescence channels
│   │   ├── C1              : nuclei          
│   │   ├── C2              : lipids
│   │   └── C3              : cytoplasm
│   └── test                : .json files for running the models in test mode for the three resolutions
│                              (across all three fluorescence channels)
├── docker                  : files for Docker image construction and running
└── exp_stats               : .csv files for image statistics and train/test splits 
    

Docker image

On AZ server:

1. Build the docker image if it doesn't exist

docker build -t 3h4m/haste .

2. Run training with docker

./run_with_docker.sh -c config.json

3. Run the shell with docker

docker run --shm-size=24gb -it --rm -v /home/group5/:/workspace 3h4m/haste bash

---

Install docker

curl -sSL https://get.docker.com/ | sh

Pull the image

sudo docker pull 3h4m/haste

Connect to the kernel

sudo docker run -it 3h4m/haste bash

---

Running inference

To run the inference you run ai_haste as a module while specifying which config file to use. All configs for inference are located in the "config/test" folder for the different magnifications. The module will run inference on all models (1 model per channel) at the same time and write the result in a pre-specified output folder.

To begin, download the models and put them in the "models" folder. Then edit each config file in "configs/test" and specify the data path and the output path.

config
│── "exp_folder":  <--- Path to output folder goes here            
│   │
│   │
│── "data"             
│   ├── "folder":"" <--- Path to data folder goes here

Inference is then executed by:

python3 -m ai_haste -c configs/test/test_20x.json 
python3 -m ai_haste -c configs/test/test_40x.json 
python3 -m ai_haste -c configs/test/test_60x.json 

Training

Training can be performed by changing selecting the data folders and magnifications and changing "run_mode" to "train" in the config and executed by

python3 -m ai_haste -c configs/train/C1/C1_20x.json 
python3 -m ai_haste -c configs/train/C2/C2_20x_pretrain.json 
python3 -m ai_haste -c configs/train/C1/C3_20x_pretrain.json 

Models

Trained models can be found here: https://drive.google.com/drive/u/0/folders/1E_Ekf-aOR8YYvodQNVXshK5WEzhSsu5k

Inference Time

Inference time can be measured in the tester/all_tester.py

in AllChannelTester.infer_all_channels() function.