VesSynth - A Robust Cross-Scale Cross-Modal 3D Vessel Segmentation Method

Warning! This repo is under active development and can change without notice. Coming soon:

• updated models
• faster code for inference
• integration into FreeSurfer!

---

Installation

1. Clone this repo and set it as your current working directory

git clone https://github.com/chiara-mauri/VesSynth.git
cd VesSynth

2. Create and activate a conda environment:

• Option 1: Use the provided yaml file:

   conda env create -f vessynth-env.yml (or for Mac: conda env create -f vessynth-env-macOS.yml)
   conda activate vessynth-env

• Option 2: Create the environment and manually install the required packages:

   conda create -n vessynth-env python=3.10
   conda activate vessynth-env
   pip install cornucopia
   pip install pandas==2.3.3 tensorstore==0.1.78 boto3

Download the models

Download the 'models' folder from https://ftp.nmr.mgh.harvard.edu/pub/dist/lcnpublic/dist/VesSynth/ unzip it and copy it in this repo

Usage

Now you can use the method with:

python path/to/repo/vessynth_test.py -i <vol> -o <outputDir> -mod <modality> [-th <threshold> -m <mask_vol> -c <cutout> -nw]

where the required arguments are:

• <vol>: input nifti volume to segment
• <outputDir>: output directory where segmentations are saved
• <modality>: modality/contrast of the input volume. Accepted values are
  • 'T2star': for exvivo MRI and all T2star-based contrasts (including SWI, QSM,...). Vessels can be both bright and dark. Mesoscopic resolution (100-400um)
  • 'HiPCT': for Hierarchical Phase-Contrast Tomography. Dark vessels. Resolution ~ 20-30um
  • 'OCT': for Optical Coherence Tomography. Dark vessels. Resolution ~ 20um
  • 'TOF': for in vivo Time-Of-Flight Magnetic Resonance angiography. Bright vessels. Flexible resolution, from ~150um iso to 500um x 500um x 1mm
  • 'fibers': for bright fiber bundles/axons across modalities (experimental)
  • 'pvs': for perivascular spaces (coming soon!)

optional arguments are:

• <threshold> value used to threshold the 'vessel probablity' to obtain a hard segmentation. Default is 0.3.
• <mask_vol> a binary mask applied to the segmentation (e.g. 1 inside brain, 0 outside). Useful to remove noise outside brain. A whole-brain mask can be obtained with mri_synthstrip
• <cutout> a bounding box to identify ROI (-zc x1 x2 y1 y2 z1 z2)
• -nw, --no_weights do NOT use Gaussian weights when computing segmentation on a patch

Synthesis pipeline

Synthetic data for training VesSynth (vessels and intensity images) have been generated using synthspline