JCell training - CTC VI Challenge

Installing the software

Install jcell and dependencies by typing bash prepare_software.sh. Alternatively, you can manually install them using the provided wheel and requirements files:

apt install wget unzip
pip install jcell-ISBI==0.0.1a4
pip install -r requirements.txt
jcell-update

2.1 Preparing Challenge datasets

Our approach operates with semantic labels, which is different from CTC challenge's instance-level annotations. For such purpose, an instance-to-semantic conversion must be applied:

python prepare_ISBI_data.py

The script will automatically download both training and test datasets. Additionally, it will slice 3D volumes converting them to 2D images. Finally, GT, ST, and GT+ST data configuration will be organized as expected by jcell.

2.2 Preparing new datasets

Prepare new datasets using the instance-to-semantic transformation script. In the following example it is shown the execution for the Fluo-C2DL-Huh7 dataset.

python prepare_new_data.py --datasets Fluo-C2DL-Huh7

The required folder structure is:

./training
├── train_val
│   └── Fluo-C2DL-Huh7
│       ├── 01
│       ├── 01_GT
│       ├── 02
│       ├── 02_GT
├── train_scripts
│   └── configuration
├── fix_relative_path.sh
├── jcell-0.0.1a4_ISBI-py3-none-any.whl
├── prepare_ISBI_data.py
├── prepare_new_data.py
├── prepare_software.sh
├── README.md
└── requirements.txt

where the folder train_val will contain a typical CTC dataset. Such a folder must be on the same level of the prepare_new_data.py script.

Configuring datasets

A json configuration file is available for specifying the path to a given dataset (./train_scripts/configuration/dataconfig_train.json). All entries must have the same structure to define a new dataset successfully. However, the current version of our software doesn't allow relatives paths. For running our current solution in your directory, we recommend executing the following script after installation:

bash fix_relative_path.sh

For adding new datasets, you can modify dataconfig_train.json directly. Please, verify that "dataset_folder" key point to your dataset.

Training with jcell

We provide all the scripts used for training. You can find them in ./train_scripts folder. For running a training job, do, for example:

cd train_scripts
bash DIC-C2DH-HeLa_GT.sh

The resulting experiment (models, logs, etc.) will be saved into ./train_scripts/out folder.

Evaluation with jcell

A general evaluation script is provided under the name general_test.sh. However, this script must be modified to properly point to the test dataset path and the corresponding model.

bash general_test.sh

NOTE: The model trained for your custom dataset will be saved into ./train_scripts/out/YOUR_DATASET/model/lastmodel.t7

fagp/ISBI-data-preparation

JCell training - CTC VI Challenge