MicroVasc-Review

Review of Segmentation and Skeletonization Methods for Large-Scale Microvascular Networks

This repository contains code and resources associated with the paper
"Segmentation and Modeling of Large-Scale Microvascular Networks: A Survey" by Goharbavang et al.

📄 Read the Full Paper

Citation

If you use this code or the paper, please cite:

@article{goharbavang2025segmentation,
  title={Segmentation and modeling of large-scale microvascular networks: a survey},
  author={Goharbavang, Helya and Ashitkov, Artem T and Pillai, Athira and Wythe, Joshua D and Chen, Guoning and Mayerich, David},
  journal={Frontiers in Bioinformatics},
  volume={5},
  pages={1645520},
  year={2025},
  publisher={Frontiers Media SA}
}

Repository Content

The paper covers the following methods:

• Binarization: Otsu’s thresholding; Frangi vesselness; Beyond Frangi; Optimally Oriented Flux; U‑Net; nnU‑Net
• Skeletonization: Lee thinning; Palágyi thinning; Surface normal accumulation (Kerautret); Fast Marching (Kline); Mean curvature flow (Tagliasacchi); Voronoi‑based (Antiga)

This repository includes two CLI modules under 'implementations/' for a subset of these methods. For the rest, please see the External Resources section below for links to community or official implementations.

• Binarize.py
  Provides otsu2d, otsu3d, frangi, and bfrangi.

• Skeleton.py
  Provides lee, palagyi, kerautret, and kline subcommands.

Other directories:

• manage_data/ Scripts for preprocessing and converting raw imaging data (e.g., Numpy → NWT).

• metrics/ Tools to compute quantitative evaluation metrics:

  • Segmentation: Jaccard index, Dice coefficient, precision, recall
  • Skeletonization: NetMets precision and recall

• optimization/ Parameter-tuning and sensitivity-analysis scripts (grid search for scale, shape, and threshold parameters).

• LICENSE MIT License.

Quick Start

1. Clone the repo

   git clone https://github.com/helia77/MicroVasc-Review.git
   cd MicroVasc-Review

2. Install dependencies (e.g., via Conda environment)

   # conda create --name microvasc
   # conda activate microvasc
   pip install -e .

3. Verify CLI tools

   binarize --help
   skeleton --help

4. Run an algorithm

   binarize frangi \
   --input path/to/volume.npy \
   --output path/to/out.npy \
   --background white \
   --params 0.5 0.5 3 \
   --scale 1 2 3 4
   --th otsu3d

External Resources & Implementations

• Beyond Frangi: https://github.com/timjerman/JermanEnhancementFilter
• Optimally Oriented Flux: https://matlab.mathworks.com/open/fileexchange/v1?id=41612
• U-Net: https://github.com/MrMras/CNN
• nnU-Net: https://github.com/MIC-DKFZ/nnunet
• Lee Thinning: https://github.com/scikit-image/scikit-image/blob/main/skimage/morphology/_skeletonize.py
• Palagyi Thinning: https://github.com/ClearAnatomics/ClearMap
• Kerautret Centerline: https://github.com/kerautret/CDCVAM
• Kline Centerline: https://github.com/TLKline/poreture
• Starlab MCF Skeletonization: https://github.com/taiya/starlab-mcfskel
• 3D Slicer: https://www.slicer.org

Refer to the paper for detailed methodology, evaluation results, and discussions.

Running External Codes

1. U-Net

• Installation:

  # Requirements
  # Python 3.12+
  # PyTorch, NumPy, OpenCV, tqdm, scipy, matplotlib
  git clone https://github.com/MrMras/CNN.git
  cd CNN
  
  conda env create -n cnn -f environment.yml
  conda activate cnn

• Usage:

    # Upload .npy files under   ./unprocessed_data/{name_of_dataset}
    # Change the path to these files in ./create_volume_from_npy.py
    python ./create_volume_from_npy.py
    # Should create a folder ./data/{name_of_dataset} and add 2 files in there.
  
    cd ./model/       # change the paths in ./model.py to ./data/{name_of_dataset}/{files}
    python ./model.py EPOCHS -c
    #  Model will save the model under ./saved_models/{name_of_dataset}/model_for_vasc3d{random id}.pth.
  
    cd ../            # specify the path for the trained model and the path to the target volume (line 64)
    python ./run_model.py
    # Binary and probability maps will be generated in ./processed_npy/ and ./probability_npy/

2. nnU-Net

• Installation:

  conda create -n nnunet python=3.9.*
  pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
  pip install nnunetv2
  
  set nnUNet_raw = path/to/nnUNet_raw
  set nnUNet_preprocessed = path/to/nnUNet_preprocessed
  set nnUNet_results = path/to/nnUNet_results

• Usage:

  # Use the new presets
  nnUNetv2_plan_experiment -d DATASET_ID -pl nnUNetPlannerResEncL
  # Extract fingerprint
  nnUNetv2_plan_and_preprocess -d DATASET_ID --verify_dataset_integrity
  # Train
  nnUNetv2_train DATASET__ID 3d_fullres 0/1/2/3/4 -p nnUNetResEncUNetLPlans
  # Predict
  nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_ID -c 3d_fullres -p nnUNetResEncUNetLPlans --save_probabilities
  nnUNetv2_apply_postprocessing -i FOLDER_WITH_PREDICTIONS -o OUTPUT_FOLDER --pp_pkl_file POSTPROCESSING_FILE -plans_json PLANS_FILE -dataset_json DATASET_JSON_FILE