VascuMap

Label-free phenotyping of human microvessel networks using deep learning segmentation and graph-based analysis.

Installation

Create and activate conda environment

conda create -n vascumap python=3.8
conda activate vascumap

Install PyTorch (adjust cuda version as needed)

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

Install the package

pip install -e .

Training

Train a new model using the following command:

python src/training.py \
    --images_dir_path "/path/to/training/images" \
    --masks_dir_path "/path/to/training/masks" \
    --model_dir_path "/path/to/save/model" \
    --model_architecture "Unet" \
    --encoder_architecture "mit_b5" \
    --input_channels 1 \
    --weights "imagenet" \
    --batch_size 16 \
    --epochs 200 \
    --learning_rate 1e-3 \
    --fp16

Training Arguments

• --images_dir_path: Directory containing training images
• --masks_dir_path: Directory containing binary mask images
• --model_dir_path: Directory to save model checkpoints
• --model_architecture: Model architecture (default: 'Unet')
• --encoder_architecture: Encoder backbone (default: 'mit_b5')
• --input_channels: Number of input channels (default: 1)
• --weights: Pretrained weights (default: 'imagenet')
• --batch_size: Batch size for training (default: 16)
• --epochs: Number of epochs to train the model (default: 200)
• --learning_rate: Learning rate for optimizer (default: 0.001)
• --fp16: Enable mixed precision training using FP16 (optional flag)

Inference

Run inference on new images using a trained model:

python src/inference.py \
    --images_dir_path "/path/to/test/images" \
    --model_checkpoint_path "/path/to/model/best.pth" \
    --output_dir_path "/path/to/output" \
    --model_architecture "Unet" \
    --encoder_architecture "mit_b5" \
    --input_channels 1 \
    --test_time_augmentation \
    --device "cuda" \
    --resolution_scale 1.0 \
    --save_prob \
    --thresholding_method "hysteresis"

Pretrained model from the paper 'Self-organization of vascularized skeletal muscle from bovine embryonic stem cells' (doi: 10.1101/2024.03.22.586252) can be found at https://zenodo.org/records/14639362

Inference Arguments

• --images_dir_path: Path to the directory containing images for inference (required)
• --model_checkpoint_path: Path to the model checkpoint (.ckpt) (required)
• --output_dir_path: Path to the directory where masks will be saved (required)
• --model_architecture: Model architecture to use for segmentation (default: 'Unet')
• --encoder_architecture: Encoder architecture to use within the model (default: 'mit_b5')
• --input_channels: Number of input channels for the model (default: 1)
• --test_time_augmentation: Enable test time augmentation (optional flag)
• --device: Device for inference, either 'cuda' or 'cpu' (default: 'cuda')
• --resolution_scale: Scale factor to apply to the image prior to inference (default: 1.0)
• --save_prob: Whether to save the probability map along with the mask (optional flag)
• --thresholding_method: Thresholding method applied to probability map (default: 'hysteresis')

Network Analysis

python src/graph_metrics.py \
--masks_dir_path "/path/to/masks/directory" \
--file_suffix "_mask" \
--save_local_metrics \
--graph_visualization

Network Analysis Arguments

• --masks_dir_path: Path to the directory containing segmentation masks (required)
• --file_suffix: Suffix of the mask files (default: '_mask')
• --save_local_metrics: Save edge and node metrics for each mask (optional flag)
• --graph_visualization: Save graph visualization (optional flag)

Data Format

• Input images should be in TIFF format
• For training, corresponding mask images should be binary (0 or 1) and in TIFF format
• Images and masks should have matching filenames

License

This work is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc/4.0/ . See the LICENSE file for details.