{
  "_ilex": {
    "architecture": "ilex.models.tract_seg.model.TractSeg",
    "constructor_kwargs": {
      "n_classes": 42,
      "n_input_channels": 9
    },
    "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",
    "weights_status": "bundled"
  },
  "authors": "Wasserthal J., Neher P. F., Maier-Hein K. H. (MIC-DKFZ, German Cancer Research Center, Heidelberg University)",
  "copyright": "TractSeg is copyright (c) MIC-DKFZ (German Cancer Research\nCenter), Apache-2.0-licensed on both the network code\n(github.com/MIC-DKFZ/TractSeg) and the pretrained weights\n(Zenodo records 3634539 + 3518348). The ilex JAX / Equinox\nport code is separately licensed under Apache-2.0 / GPL-3.0.\n",
  "data_type": "numpy",
  "description": "JAX / Equinox port of TractSeg (Wasserthal et al., *NeuroImage*\n2018; github.com/MIC-DKFZ/TractSeg). A 2D U-Net with deep\nsupervision that, per slice, predicts binary masks for each\nwhite-matter bundle (or each bundle's endpoint regions) from a\n9-channel FOD-peak input volume. The upstream's canonical\ninference path runs the per-slice forward on all three principal\nslice orientations (axial, coronal, sagittal) and averages the\nthree resulting 3D mask predictions; the ilex\nTractSegPipeline class wraps that end to end.\n\nTwo task variants ship in v0:\n- tract_segmentation_v3 (42 bundles, the upstream's current\n  default)\n- endings_segmentation_v4 (72 endpoint regions)\n\nBoth share parameter-identical bodies (~37 M trainable\nscalars); only the three 1x1 head Conv2ds at output_2,\noutput_3, and conv_5 differ in ``out_channels`` to match the\ntask's ``n_classes``.\n",
  "equinox_version": "0.13.8",
  "ilex_version": "0.0.0.dev0",
  "image_classes": "Input is a 9-channel volume ``(9, D, H, W)``, channels-first.\nThe 9 channels are 3 FOD peaks x 3 spatial components (x, y, z)\n-- the upstream's standard input for tract segmentation\nderived from a CSD-based FOD computation (MRtrix3 ``dwi2fod`` +\n``sh2peaks``).\n\nThe per-slice 2D forward consumes a single slice at a time\nalong any of the three principal axes (axial, coronal,\nsagittal); the upstream's canonical resolution is 144x144 per\nslice with 144 slices total (a 144^3 isotropic atlas grid).\n",
  "intended_use": "Research / preprocessing use for white-matter tractography\npipelines. Takes a 3D FOD peak volume (9 channels: 3 peaks x 3\nspatial components) and produces per-bundle binary masks\n(tract_segmentation) or per-bundle endpoint region masks\n(endings_segmentation). The masks then drive downstream\ntractography filtering, tract-of-interest analysis, or\nbundle-specific statistics.\n\nThe ilex port v0 wraps the raw per-slice JAX forward and the\n3-axis ensemble pipeline; consumers are responsible for the\nupstream's FOD computation (typically MRtrix3 ``dwi2fod`` +\npeak extraction), the atlas resampling to the published input\ngrid, and the downstream tractography invocation.\n",
  "jax_version": "0.10.0",
  "network_data_format": {
    "inputs": {},
    "outputs": {}
  },
  "numpy_version": "2.4.4",
  "pred_classes": "Output is the per-slice 2D logits ``(n_classes, H, W)``:\n\n* tract_segmentation_v3: ``n_classes=42`` -- one channel per\n  white-matter bundle (the upstream's refined 42-bundle set,\n  superseding the older 72-bundle v1 release).\n* endings_segmentation_v4: ``n_classes=72`` -- two endpoint\n  regions per bundle for 36 bundles, totalling 72 channels.\n\nThe TractSegPipeline class runs the forward on all three slice\norientations of a 3D volume and ensemble-averages the three\n3D mask predictions to produce the final per-voxel\nper-bundle / per-endpoint sigmoid probability volume.\n",
  "references": [
    "Wasserthal J., Neher P. F., Maier-Hein K. H. (2018). TractSeg -- Fast and accurate white matter tract segmentation. NeuroImage 183, 239-253. doi 10.1016/j.neuroimage.2018.07.070.",
    "Wasserthal J., Neher P. F., Hirjak D., Maier-Hein K. H. (2020). Multiparametric mapping of white matter microstructure in 100,000 subjects from the UK Biobank.",
    "Upstream code, license, training scripts -- github.com/MIC-DKFZ/TractSeg (Apache-2.0).",
    "Upstream weights -- Zenodo records 3634539 (tract_segmentation v3, epoch 266) and 3518348 (endings_segmentation v4, epoch 220)."
  ],
  "schema": "https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/meta_schema_20240725.json",
  "task": "white matter bundle segmentation (per-slice 2D U-Net with deep supervision; the upstream pipeline ensembles over 3 slice orientations)",
  "version": "0.0.0"
}