About the pipeline
A complete, reproducible workflow for cortical sulcal analysis
The Champollion Pipeline orchestrates the full journey from raw anatomical MRI data to compact embeddings of sulcal morphology. It leverages Morphologist (part of BrainVISA 6.0) to extract sulcal fold graphs, then uses cortical_tiles to tile 28 sulcal regions, and finally generates embeddings using pre-trained Champollion V1 transformer models hosted on Hugging Face.
A live demo is available on Hugging Face Spaces for quick single-subject testing. For batch processing or production use, install locally to leverage all available CPUs and GPUs. BrainVISA 6.0, DISCO, and Champollion are all available for download together.
Pipeline steps
Seven stages, from installation to visualisation
pixi run install-all. Pixi initialises git submodules (champollion_V1, cortical_tiles), installs all Python packages in editable mode, and creates the data directory automatically.morphologist-cli on T1 NIfTI files to extract sulcal fold graphs using Morphologist (BrainVISA 6.0). Supports serial and parallel execution via soma-workflow with a configurable CPU count.--external-config flag.full_embeddings.csv per region. Models load from local paths, archives, or directly from the Hugging Face hub (neurospin/Champollion_V1). GPU and CPU modes supported.put_together_embeddings.py. Verify with a quick ls … | wc -l check (expected: 56 files).BrainVISA ecosystem
Champollion builds on the following BrainVISA tools — part of the broader neuroimaging platform
🧠 Morphologist
Brain segmentation and sulcal morphometry. Extracts the fold graphs that feed into this pipeline. Available in BrainVISA 6.0.
🔬 Anatomist
Interactive 3D neuroimaging data visualisation. Used in step 7 to render sulcal graph meshes and cortical tile masks.
🗺 Cortical Surface
Cortical surface analysis toolbox within the BrainVISA framework, complementing sulcal region extraction.
✨ Champollion (BrainVISA)
Deep learning model for cortical folding — listed in the BrainVISA tools catalogue. This pipeline is its production deployment.
🔄 DISCO
Constraints-based diffeomorphic registration, also part of BrainVISA 6.0 and downloadable alongside Champollion.
⬇️ Download BrainVISA 6.0
BrainVISA 6.0, DISCO, and Champollion are all available in a single download from brainvisa.info.
Tutorials
Step-by-step Jupyter notebooks from the cortical_tiles and Champollion V1 repositories
The following notebooks live in the
cortical_tiles
and champollion_V1
repositories — installed locally as git submodules under external/.
Click any card to open the notebook on GitHub.
cortical_tiles — data preparation & region extraction
Compute bounding box
Derive the region-of-interest bounding box used to crop sulcal skeleton volumes.
cortical_tilesCompute mask
Generate binary sulcal masks for each of the 28 cortical regions.
cortical_tilesGenerate crops
Run the full crop-generation pipeline and inspect the resulting tile volumes.
cortical_tilesVisualise crops & masks
Overlay crop volumes on sulcal masks to verify spatial coverage.
cortical_tilesPoint cloud pattern mining
Explore sulcal fold patterns using point-cloud representations.
cortical_tilesBrainVISA preprocessing
Run BrainVISA-specific preprocessing steps before crop extraction.
cortical_tilesFoldlabel image analyses
Analyse labelled fold images: statistics, visualisation, and quality checks.
cortical_tilesChampollion V1 — embedding analysis & visualisation
Visualise augmentations
Inspect the data augmentation pipeline applied during contrastive training.
champollion_V1UMAP — UKBioBank
Project 42k UKBioBank subject embeddings onto a 2D UMAP and colour by phenotype.
champollion_V1UMAP — HCP
Explore the embedding space using HCP subjects with rich phenotypic labels.
champollion_V1Latent space & prediction (PCS)
Analyse the latent representation and downstream prediction performance for the PCS region.
champollion_V1Embedding analysis across labels
Compare how well embeddings separate different demographic and clinical labels.
champollion_V1Schizophrenia results
Visualise embedding-based predictions for a schizophrenia cohort.
champollion_V1Monozygotic twin analysis
Test whether embeddings place identical twins closer than unrelated pairs in HCP data.
champollion_V1Installation
Requires Pixi and Git — all other dependencies managed automatically
# 1. Clone the repository mkdir Champollion && cd Champollion git clone https://github.com/neurospin/champollion_pipeline.git cd champollion_pipeline # 2. Install all dependencies (submodules + Python packages) pixi run install-all # 3. Activate the Pixi environment pixi shell # 4. (Optional) Run the test suite pixi run test
Prerequisites
- Pixi package manager
- Git (with submodule support)
- BrainVISA 6.0 (for Morphologist)
- Python ≥ 3.8
Key dependencies
- PyTorch
- huggingface-hub & transformers
- Morphologist, Anatomist
- cortical_tiles, champollion_V1