You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
The Pediatric Auto-Defacer is distributed under a research-
only-restricted BSD-2-Clause license (Children's Hospital of
Philadelphia, 2024). Access is granted solely for non-commercial
academic research and educational use. Commercial use --
including clinical decision support, clinical workflows,
commercial products, or services for a fee -- is expressly
reserved by CHOP and requires separate written authorization
from Dr. Ariana Familiar (arianafamiliar@gmail.com).
Acceptance is per-repository on HuggingFace Hub; HF does not
extend acceptance across all five fold bundles automatically.
The terms below are identical for every fold in the
ilex-hub/pediatric_auto_defacer.* family: if you have already
accepted them for one fold, the same answers apply to the rest.
By requesting access you affirm that your intended use falls
within the Research-Only terms and that you have read the
license linked from this model card.
Log in or Sign Up to review the conditions and access this model content.
Pediatric MRI auto-defacer (d3b nnU-Net 5-fold ensemble) -- Pediatric Auto-Defacer fold 2 (nnU-Net V1 Generic_UNet)
Description
Pediatric Auto-Defacer (Familiar et al., AJNR 2024), ported to JAX / Equinox from the d3b-center's PyTorch nnU-Net V1 release (github.com/d3b-center/pediatric-auto-defacer-public). A five-fold ensemble of Generic_UNet (nnU-Net V1 Task070_autosegm, configuration 3d_fullres, trainer nnUNetTrainerV2, plans nnUNetPlansv2.1) that predicts a per-voxel face mask on a single-channel 3D brain MRI, then zeroes the face region in the input to produce a defaced volume. Trained on 976 multiparametric MRIs (T1w / T1w-CE / T2w / T2w-FLAIR) from 208 pediatric brain-tumour patients (Children's Brain Tumor Network) and 36 clinical controls. The network is single-channel modality-agnostic -- each modality is processed independently, NOT concatenated -- so one bundle handles all four modalities at inference time. Output: a 2-channel softmax over (background, face); the inference pipeline thresholds the face channel and applies it to the input.
Intended use
Fold 2 of the 5-fold ensemble.
Usage
from ilex.models.pediatric_auto_defacer import PediatricAutoDefacer
model = PediatricAutoDefacer.from_pretrained('ilex-hub/pediatric_auto_defacer.fold2.1')
Authors
Familiar A. M., Khalili N., Khalili N., Schuman C., Grove E., Viswanathan K., et al. (D3b Center for Data-Driven Discovery in Biomedicine, Children's Hospital of Philadelphia)
Citation
Familiar A. M., Khalili N., Khalili N., Schuman C., Grove E., Viswanathan K., Nabavizadeh A. (2024). Empowering Data Sharing in Neuroscience: A Deep Learning De-identification Method for Pediatric Brain MRIs. American Journal of Neuroradiology. doi:10.3174/ajnr.A8581.
References
- Familiar A. M., Khalili N., Khalili N., Schuman C., Grove E., Viswanathan K., Nabavizadeh A. (2024). Empowering Data Sharing in Neuroscience: A Deep Learning De-identification Method for Pediatric Brain MRIs. American Journal of Neuroradiology. doi:10.3174/ajnr.A8581.
- Upstream code + weights: github.com/d3b-center/pediatric-auto-defacer-public (Docker image
afam00/peds-brain-auto-deface; weights via Google Drive at drive.google.com/file/d/1P06VrdaMxX_VENOYVRyvFgJN82SMEsMz). - Architecture lineage: Isensee F., Jaeger P., Kohl S., Petersen J., Maier-Hein K. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18, 203-211.
License
HF Hub license tag: other
HF Hub license slug: pediatric-auto-defacer-research-only
Effective terms: Pediatric Auto-Defacer Research-Only License (c) 2024 The Children's Hospital of Philadelphia. Use of this software and the published weights is available to academic and non-profit institutions for research purposes only, subject to the terms of the 2-Clause BSD License. Commercial use -- including commercial products, services for a fee, clinical decision support, and clinical workflows -- requires direct authorization by contacting Dr. Ariana Familiar (arianafamiliar@gmail.com). The software is provided AS IS without warranty of any kind; all implied warranties of merchantability and fitness for a particular purpose are disclaimed. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.
Copyright
Network architecture and pretrained weights -- copyright (c) 2024 The Children's Hospital of Philadelphia, released under a research-only-restricted BSD-2-Clause license. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.
Upstream source
Original weights / reference implementation: https://github.com/d3b-center/pediatric-auto-defacer-public
Provenance
This artefact was produced by ilex's
save/load pipeline. The architecture is implemented in
ilex.models.pediatric_auto_defacer.PediatricAutoDefacer and the weights have been converted
from their upstream format. See the upstream source above
for the canonical reference.
- Downloads last month
- -