README.md

---
library_name: ilex
tags:
- jax
- equinox
- ilex
- neuroimaging
- 3d
license: other
license_name: fastcsr-research-only
license_link: https://github.com/IndiLab/FastCSR
gated: auto
extra_gated_prompt: |
  The FastCSR pretrained weights are distributed by the upstream
  authors (IndiLab/FastCSR) under a non-OSI Research-Only notice
  (academic research use is free; commercial use requires separate
  authorization from the upstream authors). The ilex bundle
  preserves these terms verbatim. Access is granted solely for
  non-commercial academic research and educational use. The nnUNet
  v1 Generic_UNet architecture used by FastCSR is itself
  Apache-2.0 (MIC-DKFZ); the JAX / Equinox re-expression in
  ilex.models.fastcsr inherits that architecture under Apache-2.0
  but does NOT override the FastCSR Research-Only terms governing
  the trained weights. Please confirm your use case is
  non-commercial academic research below.
extra_gated_fields:
  Name: text
  Affiliation: text
  Email: text
  Intended use: text
  I confirm non-commercial academic research use only: checkbox
extra_gated_button_content: Acknowledge and download
---

# FastCSR (cortical-surface level-set regression) -- Right-hemisphere level-set regression

## Description

FastCSR is a fast cortical-surface reconstruction pipeline that replaces FreeSurfer's implicit-surface step with a learned nnUNet v1 regression network. Three published checkpoints share the same Generic_UNet architecture (5 pool stages, 32 base features, InstanceNorm3d + LeakyReLU(0.01), convolutional pooling + ConvTranspose3d upsampling, deep_supervision=False, num_classes=1 regression head); they differ only in num_input_channels (2 for the level-set lh / rh nets that take orig + filled volumes, 1 for the brain-final-surfaces net that takes orig only). The output is a per-voxel float field representing a signed-distance level set; the surface mesh is extracted downstream by topology correction + marching cubes (nighres / JCC bindings), which is NOT vendored in v0 -- the ilex bundle ships the network forward, the level-set output is the consumer artefact. The architecture is identical to the nnUNet v1 Generic_UNet documented in Isensee et al. 2021; FastCSR's contribution is the level-set-regression training objective and the FreeSurfer-input preprocessing pipeline (orig.mgz + filled.mgz from FreeSurfer's recon-all early-stage outputs).

## Intended use

Per-voxel signed-distance level-set regression for the right-hemisphere white-matter surface. Same architecture and input contract as the lh variant.

## Usage

```python
from ilex.models.fastcsr import FastCSR
model = FastCSR.from_pretrained('ilex-hub/fastcsr.rh.1')
```

## Authors

Ren J., Hu Q., Wang W., Zhang W., Hubbard C. S., Zhang P., An N., Zhou Y., Dahmani L., Wang D., Fu X., Sun Z., Wang Y., Wang R., Li L., Liu H.

## Citation

Ren J., Hu Q., Wang W., Zhang W., Hubbard C. S., Zhang P., An N., Zhou Y., Dahmani L., Wang D., Fu X., Sun Z., Wang Y., Wang R., Li L., Liu H. (2022). Fast cortical surface reconstruction from MRI using deep learning. Brain Informatics 9(1):6. doi:10.1186/s40708-022-00155-7

### References

- Ren J., Hu Q., Wang W., Zhang W., Hubbard C. S., Zhang P., An N., Zhou Y., Dahmani L., Wang D., Fu X., Sun Z., Wang Y., Wang R., Li L., Liu H. (2022). Fast cortical surface reconstruction from MRI using deep learning. Brain Informatics 9(1):6. doi:10.1186/s40708-022-00155-7
- Isensee F., Jaeger P. F., Kohl S. A., Petersen J., Maier-Hein K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2):203-211. doi:10.1038/s41592-020-01008-z (the nnUNet v1 Generic_UNet architecture FastCSR pins to)

## License

HF Hub license tag: `other`
HF Hub license slug: `fastcsr-research-only`

**Effective terms:** FastCSR Research-Only License (per the upstream README at https://github.com/IndiLab/FastCSR). Academic research use is free; commercial use of the pretrained weights requires separate authorization from the upstream authors. The nnUNet v1 Generic_UNet architecture (vendored at staging time from https://github.com/MIC-DKFZ/nnUNet/tree/nnunetv1) is Apache-2.0 and remains so. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0 for the port code itself, but the ilex license does NOT override the upstream Research-Only terms governing the pretrained weights; non-commercial use only continues to apply.

Upstream license reference: https://github.com/IndiLab/FastCSR

### Copyright

Network architecture comes from the nnU-Net v1 codebase (DKFZ, Apache-2.0). The FastCSR-specific training code is at https://github.com/IndiLab/FastCSR (alternate mirror at https://github.com/pBFSLab/FastCSR via DeepPrep). The pretrained weights distributed by the FastCSR authors carry a non-OSI "academic research free, commercial requires authorization" notice -- not a standard open-source license; commercial use requires separate authorization from the upstream authors. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0 for the port code itself, but the ilex license does NOT override or re-license the upstream non-commercial terms governing the pretrained weights.

## Upstream source

Original weights / reference implementation: https://github.com/IndiLab/FastCSR

## Provenance

This artefact was produced by [ilex](https://github.com/hypercoil/ilex)'s
save/load pipeline. The architecture is implemented in
`ilex.models.fastcsr.FastCSR` and the weights have been converted
from their upstream format. See the upstream source above
for the canonical reference.