{
  "_ilex": {
    "architecture": "ilex.models.bme_x.model.BMEX",
    "constructor_kwargs": {},
    "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": true,
    "origin": "ilex-native",
    "weights_status": "bundled"
  },
  "authors": "Sun Y., Wang L., et al. (DBC Lab, University of North Carolina at Chapel Hill)",
  "copyright": "Network architecture and pretrained weights -- copyright (c) the DBC-Lab BME-X authors, released under the MIT License. JAX / Equinox port code -- copyright (c) the ilex authors, released under the Apache-2.0 / GPL-3.0 dual license used by ilex itself.",
  "data_type": "nibabel",
  "description": "BME-X (Sun et al., Nature Biomedical Engineering 2025), ported to JAX / Equinox from the DBC-Lab PyTorch release (github.com/DBC-Lab/Brain_MRI_Enhancement). A 3-level dual-head dense U-Net (\"Anatomy-Guided Densely-Connected U-Net\", ADU-Net) that runs a tissue classifier and a reconstruction head in parallel, bridged by a connection block that feeds the seg-derived tissue probability map back into the reconstruction pathway. The network is designed for motion correction, resolution enhancement, denoising, harmonization (cross-scanner), and 3T -> 7T-like image generation in one forward pass. Inputs are single-channel T1- or T2-weighted brain MRI patches at 32^3 voxels (the upstream's sliding-window inference tiles 64^3 patches with 0.85 overlap; the network is fully convolutional over any multiple-of-8 spatial grid). Outputs are (a) a 4-channel seg head (per-voxel relu output over BG / WM / GM / CSF; apply softmax for probabilities) and (b) a 1-channel reconstructed/enhanced volume. The published release has 15 age-and-modality-specific weight files spanning fetal-T2 through Month-24-T1/T2; per the paper, Month24 weights are also used for adult inference.",
  "equinox_version": "0.13.8",
  "ilex_version": "0.0.0.dev0",
  "image_classes": "Single-channel 3D brain MRI (T1- or T2-weighted, per the variant's modality), resampled to 0.8 mm isotropic resolution per the upstream preprocessing, patched into 32^3 or 64^3 windows for inference.",
  "intended_use": "Inference-only enhancement of a single-channel 3D brain MRI volume. Inputs are expected to be conformed to 0.8 mm isotropic, brain- extracted, and intensity-normalised via histogram-matching against an age-specific template (the upstream's BME_X_enhanced.py pipeline performs all three steps), then tiled into 32^3 patches with sliding-window inference at 0.85 overlap. Each variant's weights are tuned for a specific age range + modality (T1 vs T2). Per Sun et al. NBME 2025, the Month24 variants are the de-facto adult-applicable models. The seg head's 4-channel pre-softmax relu output (BG / WM / GM / CSF; apply softmax for probabilities) is an auxiliary signal that drives the reconstruction head's tissue-aware enhancement; the recon head's 1-channel output is the enhanced volume (motion- corrected, denoised, super-resolved, harmonized in one forward). IMPORTANT input contract -- BME-X expects input at the cohort- specific template's raw intensity scale (0-1000), NOT naively renormalised to [0, 1]. The v0.1 pipeline (ilex.models.bme_x.pipeline.enhance) handles this: it fetches the cohort template lazily from OSF via lytemaps.fetch_bmex, histogram- matches the input, tiles into 32^3 patches with 0.85 overlap, and aggregates the recon outputs. Consuming BMEX.__call__ directly on naively-normalised input produces all-zero recon output for about half the variants (including both adult-applicable ones) -- this is the input-contract mismatch, not a broken model (PyTorch and JAX agree to ~5e-5 in all measured cases). End-to-end inference on a real adult T1 patch (cvs_avg35 MNI, central 64^3) through pipeline.enhance with cohort='adult', suffix='T1w' produces recon range [0, 533.6] with 58.8% nonzero voxels and a diverse 4-class tissue segmentation; the pipeline is the validated usability surface.",
  "jax_version": "0.10.0",
  "network_data_format": {
    "inputs": {},
    "outputs": {}
  },
  "numpy_version": "2.4.4",
  "pred_classes": "Two heads. (1) seg -- 4-channel per-voxel relu output over BG / WM / GM / CSF; apply softmax(axis=0) for tissue probabilities. (2) recon -- 1-channel enhanced volume at the input spatial resolution.",
  "references": [
    "Sun Y., Wang L., Li G., Lin W., Liu M., Wang Y., He X., Wang Z., Wei Z., Han Z., Tian Y., Niu J., Cui Z., Zhu D., Yang Y., Zhao L., Wei Z., Quinn B. T., Schiratti J.-B., Zhu Y., Wei H., Wang W., Tao Y., Zhang Y., Wang Z., Cui Z., Bashyam V., Wen J., Tassinari G., Davatzikos C., Mowery D. L., Zhou Z., Yan J., Sun L., Shen D., Wang L. (2025). A foundation model for enhancing magnetic resonance images and downstream segmentation, registration and diagnostic tasks. Nature Biomedical Engineering. doi:10.1038/s41551-024-01283-7.",
    "Upstream code: github.com/DBC-Lab/Brain_MRI_Enhancement (BME_X/models/DUNet3D_seg_recon_softmax.py)."
  ],
  "schema": "https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/meta_schema_20240725.json",
  "task": "Brain MRI enhancement (tissue-aware foundation model)",
  "version": "0.0.0"
}