You can deploy this model and start earning money today!

AishwariyaDutta/
pvs-seg-nnunet-checkpoints

Name: pvs-seg-nnunet-checkpoints
Availability: InStock

Views

No views yet

PVS-Seg — Perivascular Space Segmentation (3T T2w MRI, nnU-Net v2)

Trained weights for segmenting perivascular spaces (PVS) in 3T T2-weighted brain MRI. This is a 5-fold nnU-Net v2 ensemble (Dataset401_PVS3T80, 3d_fullres).

These are weights only. For an installable inference package, use pvs-seg-nnunet on PyPI, which downloads these checkpoints automatically:

bash

pip install pvs-seg-nnunet
pvsseg predict -i /path/to/inputs -o /path/to/outputs

Model


Architecture	nnU-Net v2, `3d_fullres`, PlainConvUNet (~30.8M params)
Trainer	`nnUNetTrainerCustomAug` (custom intensity augmentation)
Input	single channel, T2w (`CASE_0000.nii.gz`)
Labels	`0 = background`, `1 = PVS`
Target spacing	0.8 × 0.8 × 0.8 mm (isotropic)
Patch size	96 × 160 × 160
Normalization	Z-score
Ensemble	5 folds (`fold_0` … `fold_4`), `checkpoint_best.pth` each

Training data

200 subjects from HCP-Aging, fixed 80/20 split (seed 42).
160 training cases (5-fold cross-validation), 40 held-out test cases.
All data acquired at 3T.

Evaluation (40-case held-out 3T test set)

Metric	Mean	Median
Dice	0.884	0.884
IoU	0.795	0.793
Precision	0.913	0.927
Recall	0.861	0.871
Surface Dice (1 mm)	0.942	0.948
HD95 (mm)	1.52	1.13

Lesion-wise (connected-component detection): F1 0.834, precision 0.865, recall 0.822. Performance is strongest on large lesions (Dice 0.870, recall 0.97) and drops for small lesions (≤10 voxels; Dice 0.674) — small-PVS detection is the known limitation of this model.

Intended use & limitations

Intended: research segmentation of PVS in 3T T2w MRI at ~0.8 mm isotropic resolution.
Out of scope / use with caution: other field strengths (e.g. 7T), other contrasts (T1w/FLAIR), substantially different resolutions or acquisition protocols, and clinical/diagnostic use. Domain shift will degrade results.
Inputs must be skull-comparable to the training distribution; follow standard nnU-Net preprocessing expectations.

Repository layout

nnUNetTrainerCustomAug/
├── plans.json
├── dataset.json
├── dataset_fingerprint.json
└── fold_{0..4}/
    └── checkpoint_best.pth   # used by the package (periodic checkpoints also present)

The repo also stores periodic training checkpoints (checkpoint_10 … checkpoint_1000). For inference you only need checkpoint_best.pth per fold plus plans.json / dataset.json; the PyPI package downloads exactly those.

How to load manually

python

import torch
from huggingface_hub import snapshot_download
from nnunetv2.inference.predict_from_raw_data import nnUNetPredictor

# requires the nnUNetTrainerCustomAug class to be importable
# (the pvs-seg-nnunet package registers it for you)
root = snapshot_download(
    "AishwariyaDutta/pvs-seg-nnunet-checkpoints",
    allow_patterns=[
        "nnUNetTrainerCustomAug/plans.json",
        "nnUNetTrainerCustomAug/dataset.json",
        "nnUNetTrainerCustomAug/dataset_fingerprint.json",
        "nnUNetTrainerCustomAug/fold_*/checkpoint_best.pth",
    ],
)
model_dir = f"{root}/nnUNetTrainerCustomAug"

predictor = nnUNetPredictor(device=torch.device("cuda"))
predictor.initialize_from_trained_model_folder(
    model_dir, use_folds=(0, 1, 2, 3, 4), checkpoint_name="checkpoint_best.pth"
)
predictor.predict_from_files("inputs/", "outputs/")

License

MIT.

AishwariyaDutta/pvs-seg-nnunet-checkpoints