ilex-hub/asl_denoising.published.1

ASL Denoising Autoencoder (Hales et al. -- ASL difference image denoising) -- ASL DAE published checkpoint

Description

patrickhales/asl-denoising (Hales et al., JMRI 2020) is a 5-Conv2D U-Net-shaped denoising autoencoder for arterial spin labelling (ASL) MRI difference images (dM). The model takes single-repetition raw dM images (high-noise, single label- control subtraction) and emits a denoised dM image approximating the multi-repetition mean (low-noise, averaged over typically 10 repetitions). The model also suppresses transient artefacts from head motion, arterial-transit variation, and spurious perfusion signals that survive a single-rep subtraction.

v0 ships one variant (the upstream's released DaeTrainedModel.h5; trained on 28,820 dM images over 67 early-stopped epochs of a planned 100). The architecture is parameter-fixed at 149,441 trainable scalars.

Intended use

ASL difference-image denoising + transient-artefact suppression. Input: (1, 128, 128) z-normalised dM slice (the consumer applies the preprocessing constants from the bundle's _ilex.preprocessing block). Output: same shape, also z-normalised; the consumer un-normalises by output * std + mean. The model is shape-invariant for any (H, W) divisible by 4; the upstream's pipeline always resizes inputs to (128, 128) first.

Usage

from ilex.models.asl_denoising import AslDenoisingDAE
model = AslDenoisingDAE.from_pretrained('ilex-hub/asl_denoising.published.1')

Authors

Hales P. W. (Great Ormond Street Hospital / UCL Institute of Child Health)

Citation

Hales P. W., Pfeuffer J., Clark C. A. (2020). Combined denoising and suppression of transient artefacts in arterial spin labelling MRI using deep learning. Journal of Magnetic Resonance Imaging. doi:10.1002/jmri.27255.

References

• Hales P. W., Pfeuffer J., Clark C. A. (2020). Combined denoising and suppression of transient artefacts in arterial spin labelling MRI using deep learning. Journal of Magnetic Resonance Imaging. doi 10.1002/jmri.27255.
• Upstream code + weights -- github.com/patrickhales/asl-denoising (GPL-3.0).

License

HF Hub license tag: gpl-3.0

Effective terms: GPL-3.0-only. The upstream code + released Keras .h5 weights are licensed under GPL-3.0 at github.com/patrickhales/asl-denoising. The ilex JAX / Equinox port code is licensed under GPL-3.0 to preserve the upstream's terms; this port is NOT available under the Apache-2.0 track that covers ilex's permissive- licensed ports. Redistribution requires the GPL-3.0 source-code-availability obligation.

Upstream license reference: https://www.gnu.org/licenses/gpl-3.0.en.html

Copyright

patrickhales/asl-denoising is copyright (c) Patrick Hales, GPL-3.0-licensed on the code + the released Keras .h5 weights. The ilex JAX / Equinox port code is licensed under GPL-3.0 to preserve the upstream's license; this port is NOT available under the Apache-2.0 track that covers ilex's permissive-licensed ports.

Upstream source

Original weights / reference implementation: https://github.com/patrickhales/asl-denoising

Provenance

This artefact was produced by ilex's save/load pipeline. The architecture is implemented in ilex.models.asl_denoising.AslDenoisingDAE and the weights have been converted from their upstream format. See the upstream source above for the canonical reference.