{
  "_ilex": {
    "architecture": "ilex.models.synthmorph_deform.model.SynthMorphDeform",
    "constructor_kwargs": {
      "add_nf": [
        256,
        256,
        256,
        256
      ],
      "dec_nf": [
        256,
        256,
        256,
        256
      ],
      "enc_nf": [
        256,
        256,
        256,
        256
      ],
      "flow_nf": 3,
      "nb_hyp_layers": 4,
      "nb_hyp_params": 1,
      "nb_hyp_units": 32,
      "src_feats": 1,
      "trg_feats": 1
    },
    "format": "ilex",
    "framework_version": {
      "equinox": "0.13.8",
      "ilex": "0.0.0.dev0",
      "jax": "0.10.0",
      "jaxlib": "0.10.0",
      "numpy": "2.4.4",
      "safetensors": "0.7.0"
    },
    "has_state": false,
    "origin": "ilex-native"
  },
  "authors": "Hoffmann M., Hoopes A., Greve D. N., Iglesias J. E., Fischl B., Dalca A. V.",
  "copyright": "Network architecture, training code, and pretrained weights: copyright (c) the SynthMorph / VoxelMorph authors and the FreeSurfer maintainers, distributed via the FreeSurfer software distribution under the FreeSurfer Software License (FSLA; permissive academic / non-commercial research use). See https://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferSoftwareLicense for the binding terms. The voxelmorph reference implementation itself is dual- Apache-2.0 / GPL-3.0; the SynthMorph weights are a downstream artefact of voxelmorph + synthetic-data training, distributed through the FreeSurfer bundle. JAX / Equinox port code: copyright (c) the ilex authors, released under the Apache-2.0 / GPL-3.0 dual license used by ilex itself; the ilex port covers only the original Equinox re-expression and does not override the upstream FreeSurfer / voxelmorph terms.",
  "data_type": "nibabel",
  "description": "SynthMorph deformable (Hoffmann et al. 2024, *Imaging Neuroscience* 2:1-33) ported to JAX / Equinox from the FreeSurfer-bundled VoxelMorph reference implementation. The network is a HyperMorph-style hypernetwork wrapped around a 5-level VxmDense U-Net -- a 4-layer dense MLP (32 units per layer) takes a single hyperparameter (the deformation regularization weight) and outputs the conv kernels + biases for all 13 conv layers of the embedded U-Net at forward time. The published ``synthmorph.deform.3.h5`` (~3.5 GB) carries the dense-projection weights that map the hypernet embedding to each conv layer's kernel + bias; the conv layers themselves hold no static weights. The v0 ilex bundle returns the raw 3-channel velocity field at the input spatial resolution; downstream integration (squaring-and-scaling VecInt) and spatial-transform warp are parameter-free pure numerics and live outside the v0 bundle, mirroring the affine port's barycenter / fit_affine deferral.",
  "equinox_version": "0.13.8",
  "ilex_version": "0.0.0.dev0",
  "image_classes": "Per-sample 3D medical-imaging volumes (moving + fixed). Intensity normalisation matches the upstream ``transform(..., normalize=True)`` step (clip to per-sample 99th percentile; rescale to ``[0, 1]``). Both volumes must be at the same spatial grid; the published checkpoint is trained at a conformed input shape and benefits from upstream-style padding when the caller's grid differs.",
  "intended_use": "Deformable registration of two 3D medical-imaging volumes, hypernetwork-conditioned on the deformation regularization weight (the user picks any value in [0, 1] at inference time without retraining). The v0 ilex bundle returns the raw velocity field; the user composes integration + spatial-transform warp externally (a ``nimox.modules`` follow-up, mirroring the affine port's barycenter / fit_affine deferral).",
  "jax_version": "0.10.0",
  "network_data_format": {
    "inputs": {},
    "outputs": {}
  },
  "numpy_version": "2.4.4",
  "pred_classes": "3-channel velocity field at the input spatial resolution. Each channel is the per-voxel displacement contribution along one spatial axis, in voxel units; the downstream integration step (squaring-and-scaling VecInt) converts the velocity field to the per-voxel diffeomorphic displacement, and the spatial-transform step then warps the source volume using the displacement field. Both of those downstream steps are parameter-free and out-of-scope at v0.",
  "references": [
    "Hoffmann M., Billot B., Greve D. N., Iglesias J. E., Fischl B., Dalca A. V. (2022). SynthMorph: learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging, 41(3):543-558. doi:10.1109/TMI.2021.3116879.",
    "Hoffmann M., Hoopes A., Greve D. N., Fischl B., Dalca A. V. (2024). Anatomy-aware and acquisition-agnostic joint registration with SynthMorph. Imaging Neuroscience, 2:1-33. doi:10.1162/imag_a_00197.",
    "Hoopes A., Hoffmann M., Greve D. N., Fischl B., Guttag J., Dalca A. V. (2022). Learning the effect of registration hyperparameters with HyperMorph. Journal of Machine Learning for Biomedical Imaging, 1:1-30."
  ],
  "schema": "https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/meta_schema_20240725.json",
  "task": "Hypernetwork-conditioned deformable image-to-image registration",
  "version": "0.0.0"
}