Version: v1.0.0

This repository releases the project-authored processing scripts and the evaluation adaptation used with the BraTS-GLI-Anatomy-Lesion derivative resource. It is distributed separately from the data package.

This repository provides:

1. A wrapper for externally supplied healthy-tissue probability-map inference.
2. Aggregation of source probabilities into the eight-class joint label space.
3. Entropy-weighted fusion with a supplied lesion-mask constraint.
4. Lesion inpainting orchestration with an externally installed NiftySeg, intended for cases marked image_repair_required=repair_required.
5. Evaluation files adapted from the MedNeXt/nnUNet evaluation package for the reported joint-label metric behavior.

This repository does not distribute BraTS imaging data, model weights, MedNeXt model or training code, DeepWMH, LST-AI, TumorSynth, NiftySeg, or a complete third-party software environment. MedNeXt was used as a validation model; this release does not modify or redistribute its model architecture, trainer, or weights.

CODE/
|-- .gitignore
|-- LICENSE.txt
|-- LICENSE-APACHE-2.0.txt
|-- README.md
|-- THIRD_PARTY_NOTICES.md
|-- requirements.txt
|-- run_inpainting.bash
|-- evaluation/
|   |-- README.md
|   |-- evaluator.py
|   `-- metrics.py
`-- healthy label/
    |-- nnunet.sh
    |-- aggregate_seg.py
    `-- shang.py

External tools and trained models must be obtained from their official sources and are not redistributed here:

• MedNeXt: https://github.com/MIC-DKFZ/MedNeXt
• NiftySeg: https://github.com/KCL-BMEIS/NiftySeg
• DeepWMH: https://github.com/lchdl/DeepWMH
• LST-AI: https://github.com/CompImg/LST-AI
• TumorSynth: https://github.com/fprados/TumorSynth

The Python packages imported by the released scripts are listed in requirements.txt. Installation and licensing of external tools and models remain governed by their official sources.

healthy label/nnunet.sh invokes an externally available nnUNet_predict command with externally supplied trained-model assets. Inputs must be named <subject>_0000.nii.gz through <subject>_0003.nii.gz.

export NNUNET_RAW_DATA_BASE=/path/to/nnUNet_raw
export NNUNET_PREPROCESSED=/path/to/nnUNet_preprocessed
export NNUNET_RESULTS_FOLDER=/path/to/nnUNet_trained_models
export NNUNET_INPUT_DIR=/path/to/imagesTr
export NNUNET_OUTPUT_DIR=/path/to/inference_outputs
bash "healthy label/nnunet.sh"

Optional variables are MODEL_NAME, TASK, MAX_JOBS, and START_INDEX. This wrapper does not include a trained model, training configuration, or model-weight download.

healthy label/aggregate_seg.py aggregates externally generated probability maps into the joint label space:

python "healthy label/aggregate_seg.py" \
  --input-dir /path/to/inference_outputs \
  --output-dir /path/to/aggregated_outputs \
  --workers 16

The script saves aggregated probability maps, NIfTI segmentations, and voxel_stats_summary.csv. The source-channel mapping encoded by this script must be used with the corresponding external source probability model.

healthy label/shang.py fuses aggregated per-modality probabilities and imposes the supplied lesion mask as class 4:

python "healthy label/shang.py" \
  --npz-root /path/to/aggregated_outputs \
  --expert-root /path/to/lesion_masks \
  --output-dir /path/to/fused_outputs \
  --workers 8

The current implementation accepts lesion-mask filenames matching NNNN.nii.gz or NNNN-seg.nii.gz, where NNNN is a four-digit case ID. It encodes modality exclusions for cases 0845, 1138, 1149, 1158, and 1182; changes to those exclusions alter this release's curation decision and must be reported.

The associated Synapse release provides metadata/id_mapping.csv, which maps these four-digit IDs back to the original BraTS-compatible case identifiers and file names for locating excluded TumorSynth-derived modality probability maps.

run_inpainting.bash orchestrates NiftySeg executables obtained separately from the official project. The script uses the pre-joint-label convention in which label 2 is inpainted and a dilated label 1 region is excluded from the brain mask. For cases marked image_repair_required=repair_required in the associated data resource, use the image-repair labels from repair_labels/modified/ as the inpainting-mask input when reconstructing the repaired MRI inputs used during curation. Expert-negative cases do not use these repair labels.

export BASE_DATA_DIR=/path/to/t1n_t1c_t2w_t2f_inputs
export LABEL_DIR=/path/to/inpainting_masks
export OUTPUT_DIR=/path/to/inpainted_outputs
bash run_inpainting.bash

Optional variables are NIFTYSEG_BIN_DIR, MAX_JOBS, DILATION, and SEARCH_SIZE.

evaluation/evaluator.py and evaluation/metrics.py are Apache-2.0-licensed adaptations of files from the MedNeXt/nnUNet evaluation package. To reproduce the joint-label evaluation used in this work, obtain and install MedNeXt from its official repository, then replace the corresponding MedNeXt evaluation files with the two adapted files released in evaluation/.

The adapted DSC and HD95 metric rules are documented in evaluation/README.md. The released evaluation files do not constitute a modification of the MedNeXt model architecture, trainer, or model weights.

Project-authored scripts and documentation are released under the MIT License in LICENSE.txt.

evaluation/evaluator.py and evaluation/metrics.py are modified from the MedNeXt/nnUNet evaluation package and are distributed under the Apache License, Version 2.0, provided in LICENSE-APACHE-2.0.txt. Their upstream attribution and modification notice are described in THIRD_PARTY_NOTICES.md.

The associated derivative data resource is licensed separately from this code package and remains subject to applicable upstream BraTS requirements. The data resource is hosted on Synapse at https://www.synapse.org/Synapse:syn75210889/wiki/.

The associated BraTS-GLI-Anatomy-Lesion v1.0.0 data resource is a labels-only Synapse resource with a public landing page and an access-controlled versioned label release:

• Public Synapse landing page and Wiki: https://www.synapse.org/Synapse:syn75210889/wiki/640990
• Synapse project ID: syn75210889
• Controlled label release: syn75260521
• Upstream BraTS 2023 project ID: syn51156910
• Upstream BraTS access page: https://www.synapse.org/Synapse:syn51156910/wiki/627000
• Code repository: https://github.com/xyx200/brats-gli-anatomy-lesion-code
• Official contact: xxy200200@stu.xjtu.edu.cn
• Citation: Until a final peer-reviewed bibliographic citation is available, cite the current resource article preprint: Xiang, X., Hao, S., Wang, F., Ma, J., and Lian, C. (2026). GLI-AL: A Multi-Modal Glioma MRI Label Resource with Unified Anatomy-Lesion Labels. arXiv:2607.22135. https://doi.org/10.48550/arXiv.2607.22135. The associated data resource may additionally be cited alongside the resource article using its DOI: https://doi.org/10.7303/SYN75210889.

The data package contains labels for 1,251 BraTS 2023-GLI cases, split into 394 purified cases and 857 extended cases. Its labels-only release manifest lists 1,367 NIfTI label files: 1,251 released unified segmentation labels and 116 image-repair labels. The image-repair labels are derived from the SVID/WMH label resources associated with: Multi-Disease Segmentation of Gliomas and White Matter Hyperintensities in the BraTS Data Using a 3D Convolutional Neural Network, and the source repository johncolby/svid_paper.

MRI images are not redistributed in this Synapse project. Users must obtain the corresponding BraTS 2023-GLI MRI data directly from upstream Synapse ID syn51156910; for the 116 cases marked image_repair_required=repair_required, users should additionally apply the image-repair labels with the provided lesion-inpainting script (run_inpainting.bash) to reconstruct the repaired MRI inputs used during curation. Expert-negative cases are marked image_repair_required=no_repair_required and do not have repair labels.

The controlled release includes the following metadata files: participants.csv, labels_description.json, metadata/cases.csv, metadata/id_mapping.csv, metadata/release_manifest.csv, metadata/data_dictionary.md, and scripts/download.sh.

The Synapse project landing page and Wiki are publicly viewable. The versioned label release (syn75260521) remains access-controlled. Access requires a registered Synapse account and approved access to the upstream BraTS 2023 data through syn51156910. Follow the email-based request instructions on the public Wiki; after manual approval, the maintainer grants the requester's Synapse account read and download access to syn75260521. The labels are released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). This data license does not remove or replace the John Colby MIT notice for the SVID/WMH repair-label source material and does not override applicable upstream BraTS data-use terms.

Downloaded label files must not be shared with anyone who has not been approved for access to the derivative label resource.