VesselBoost pretrained weights

Model purpose

VesselBoost segments small blood vessels in high-resolution human brain MRI. The primary models target time-of-flight magnetic resonance angiography (TOF-MRA). One checkpoint, t2s_mod_ep1k2_0728, provides experimental support for T2*-weighted MRI.

These files are PyTorch state dictionaries for use with the VesselBoost inference, test-time adaptation, and boosting workflows. They are not standalone Hugging Face Transformers models or hosted inference endpoints.

Research use only. Not validated for clinical diagnosis, treatment planning, or other clinical decision-making.

Architecture and release pin

The checkpoints use the VesselBoost 3D U-Net with one input channel, one output channel, and 16 base filters. The network has four encoder stages, a bridge, four decoder stages with transposed-convolution upsampling and skip connections, and a final 1 x 1 x 1 convolution. Each convolutional block contains two 3 x 3 x 3 convolutions with batch normalization and ReLU activation.

The corresponding source release is pinned to:

See config.json for the machine-readable inference configuration.

Checkpoints

All pretrained checkpoints are stored under weights/. Their original filenames and serialization formats are preserved from the original release.

Checkpoint MRI contrast Description
BM_VB2_aug_all_ep2k_bat_10_0903 TOF-MRA Primary TOF-MRA checkpoint referenced by the VesselBoost documentation and tests; trained with the combined augmentation configuration.
VB2_aug_intensity_ep2k_bat10_0903 TOF-MRA Augmentation ablation using the intensity augmentation configuration.
VB2_aug_off_ep2k_bat10_0903 TOF-MRA Augmentation ablation with augmentation disabled.
VB2_aug_random_ep2k_bat10_0903 TOF-MRA Augmentation ablation using the random augmentation configuration.
VB2_aug_spatial_ep2k_bat10_0903 TOF-MRA Augmentation ablation using the spatial augmentation configuration.
manual_0429 TOF-MRA Legacy checkpoint associated with the manual-label training run and used in VesselBoost v2.0.2 examples.
omelette1_0429 TOF-MRA Legacy TOF-MRA checkpoint identified as Omelette variant 1.
omelette2_0429 TOF-MRA Legacy TOF-MRA checkpoint identified as Omelette variant 2.
t2s_mod_ep1k2_0728 T2*-weighted MRI Experimental T2*-weighted vessel-segmentation checkpoint. It has not received the same validation as the primary TOF-MRA model.

The augmentation-specific checkpoints are included to preserve the original model set and support comparison or reproduction of augmentation experiments. For the standard TOF-MRA prediction workflow, use manual_0429 or BM_VB2_aug_all_ep2k_bat_10_0903 unless reproducing a specific legacy experiment.

Downloading checkpoints

Install the Hugging Face command-line client:

python -m pip install huggingface_hub

Download the primary TOF-MRA checkpoint:

hf download BrainVascuLab/VesselBoost \
  weights/BM_VB2_aug_all_ep2k_bat_10_0903 \
  --local-dir saved_models

The downloaded checkpoint will be available at saved_models/weights/BM_VB2_aug_all_ep2k_bat_10_0903.

Download every pretrained checkpoint and the checksum manifest:

hf download BrainVascuLab/VesselBoost \
  --include "weights/*" \
  --local-dir saved_models

For reproducible automated workflows, pass --revision with a specific Hugging Face commit hash rather than relying on the moving main branch.

Preprocessing and inference

VesselBoost v2.0.2 performs the following inference operations:

  1. Load a single-channel NIfTI MRI volume.
  2. Resize each spatial dimension to at least 64 voxels and to a multiple of 64, using nearest-neighbor interpolation.
  3. Apply whole-volume z-score standardization: subtract the volume mean and divide by its standard deviation. A constant-valued volume is mapped to zeros.
  4. Divide the standardized image into non-overlapping 64 x 64 x 64 patches. The optional Gaussian-blending path uses overlapping patches.
  5. Apply the 3D U-Net and a sigmoid activation to obtain vessel probabilities.
  6. Threshold probabilities at the default value of 0.1.
  7. Remove connected components smaller than 10 voxels using 26-connectivity.
  8. Resize the prediction back to the original image dimensions.

VesselBoost preprocessing modes can optionally perform N4 bias-field correction, denoising, both operations, or neither. Use the same preprocessing choices used for validation when comparing results. Brain extraction is optional and requires separate SynthStrip weights; those third-party weights are not part of this model release.

Integrity verification

SHA-256 checksums for every checkpoint are provided in weights/MANIFEST.sha256. After downloading all files, verify them with:

cd saved_models/weights
sha256sum --check MANIFEST.sha256

All nine checkpoints should report OK.

Load the checkpoints with the pinned VesselBoost source and map tensors to the intended device. When supported by the installed PyTorch version, use weights_only=True when loading these state dictionaries.

Known limitations and expected failure cases

  • The models were developed for research MRI data and may not generalize to unseen scanners, field strengths, acquisition protocols, resolutions, populations, pathologies, or non-brain anatomy.
  • The primary models target TOF-MRA. Applying them to other contrasts can produce unreliable results; T2* support is explicitly experimental.
  • Bright non-vascular structures, noise, motion, ringing, bias fields, susceptibility artifacts, and incomplete brain masking can cause false positives.
  • Low vessel contrast, slow or turbulent flow, signal dropout, very small vessels, severe pathology, and partial-volume effects can cause false negatives or disconnected vessels.
  • Z-score standardization is performed over the supplied volume. Large background regions, unexpected cropping, NaN or infinite intensities, and constant-valued images can change or invalidate the result.
  • Resizing and patch boundaries can alter fine structures. Gaussian blending may reduce patch-boundary artifacts but changes the inference procedure and should be reported.
  • The default probability threshold of 0.1 and component cutoff of 10 voxels may require validation for a new dataset. Tuning them on evaluation cases can bias reported performance.
  • Training labels for small vessels can be incomplete or imperfect. Predictions should not be interpreted as a complete representation of the cerebral vasculature.
  • Detailed provenance for manual_0429, omelette1_0429, and omelette2_0429 training runs is documented in our ApertureNeuro journal article VesselBoost: A Python Toolbox for Small Blood Vessel Segmentation in Human Magnetic Resonance Angiography Data.

GitHub Actions CI outputs

The latest generated outputs from VesselBoost's GitHub Actions test workflows are stored in the public VesselBoost CI bucket.

The bucket uses the following layout:

github_actions/
β”œβ”€β”€ boost/predicted_labels/
β”œβ”€β”€ docker/saved_model/
β”œβ”€β”€ prediction/predicted_labels/
β”œβ”€β”€ train/saved_model/
└── tta/predicted_labels/

These files are automated CI diagnostics, not validated model releases or benchmark results. Each successful push-triggered workflow replaces the previous contents of its corresponding directory.

Resources and citation

Please cite:

@article{xuVesselBoostPythonToolbox2024,
  title = {VesselBoost: A Python Toolbox for Small Blood Vessel Segmentation in Human Magnetic Resonance Angiography Data},
  author = {Xu, Marshall and Ribeiro, Fernanda L. and Barth, Markus and Bernier, Micha\"el and Bollmann, Steffen and Chatterjee, Soumick and Cognolato, Francesco and Gulban, Omer F. and Itkyal, Vaibhavi and Liu, Siyu and Mattern, Hendrik and Polimeni, Jonathan R. and Shaw, Thomas B. and Speck, Oliver and Bollmann, Saskia},
  journal = {Aperture Neuro},
  volume = {4},
  year = {2024},
  doi = {10.52294/001c.123217}
}

License

The files in this model release are provided under the MIT License. See LICENSE.

Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support