EMtrain

Pipeline for training convolutional neural networks to segment neurons from volume electron microscopy data. Based on local-shape descriptors.

Installation

pip install -e .

Or install dependencies directly:

pip install -r requirements.txt

Usage

Training

Start a new training run:

python emtrain/train.py -p <PROJECT_DIR> -cfg <CONFIG_JSON> [--gpu-id 0] [-c NUM_WORKERS]

Resume an existing experiment:

python emtrain/train.py -p <PROJECT_DIR> -cfg <CONFIG_JSON> -r <EXPERIMENT_ID>

Options:

• -p, --project-dir: Path to the project directory
• -cfg, --config: Path to training configuration JSON
• --gpu-id: GPU device ID (default: 0)
• -c, --cache-workers: Number of cache workers for data loading
• -r, --resume: Experiment ID to resume
• --no-comet-log: Disable Comet ML logging

Model Evaluation

Evaluate trained models against ground truth annotations:

python evaluate_models.py

Configuration

Training and evaluation are controlled via JSON configuration files. Examples are provided in:

• training_config.json - Training parameters
• ground_truth_config.json - Ground truth dataset paths
• seg_config.json - Segmentation pipeline configuration
• volumes config JSONs - Per-volume evaluation configurations

Models

Trained models are available at: https://github.com/stanleyheinze/auto-segmentation-models

Author

Valentin Gillet (valentin.gillet@biol.lu.se)

License

MIT License.

Acknowledgments

EMtrain is built on gunpowder and local-shape descriptors by the Funke lab. Also see the publication associated with local-shape descriptors: Local shape descriptors for neuron segmentation