broadinstitute/axonet-neuromorpho-dataset

AxoNet NeuroMorpho Dataset

Multi-view rendered images of neuronal morphologies from NeuroMorpho.org, prepared for training segmentation and multimodal (CLIP-style) models.

Dataset Summary

This dataset contains 7,158 curated neurons rendered from 24 viewpoints each, totaling approximately 164,000 images. Each neuron has:

• Binary masks (silhouette)
• Semantic segmentation masks (soma, axon, dendrites)
• Depth maps
• Rich metadata (species, brain region, cell type, morphometrics)

Supported Tasks

• Semantic Segmentation: Predict neuron compartment labels (soma, axon, basal/apical dendrites)
• Depth Estimation: Predict relative depth from 2D projections
• Contrastive Learning: Train CLIP-style models for text-image retrieval
• Zero-shot Classification: Classify neurons by cell type or brain region using natural language

Dataset Structure

data/
  curated_manifest.jsonl   # 7,158 neurons (QC'd, species-balanced)
  full_manifest.jsonl      # 164,016 images (all views)
  metadata.jsonl           # NeuroMorpho metadata per neuron

images/
  [neuron_id]/
    [neuron_id]_0000_mask_bw.png     # Binary mask
    [neuron_id]_0000_mask.png        # Semantic segmentation
    [neuron_id]_0000_mask_color.png  # Color-coded visualization
    [neuron_id]_0000_depth.png       # Depth map
    ... (24 views per neuron)

provenance/
  curation_report.txt      # QC statistics
  download_log.jsonl       # Download metadata
  render_config.json       # Rendering parameters

Data Fields

Curated Manifest Fields (curated_manifest.jsonl)

One record per neuron (7,158 total):

• neuron_id: Unique NeuroMorpho.org identifier (integer)
• neuron_name: NeuroMorpho.org name string
• swc: Path to SWC morphology file
• species: Species name
• brain_region: List of anatomical regions
• cell_type: List of cell type classifications
• archive: Source archive/lab
• physical_Integrity: Data quality annotation

Full Manifest Fields (full_manifest.jsonl)

One record per rendered view (~164K total):

• neuron_id: Neuron identifier (string)
• swc: SWC filename
• mask: Path to semantic segmentation mask
• mask_bw: Path to binary mask
• depth: Path to depth map
• idx: View index (0-23)
• camera: Camera parameters (eye, target, up, fovy, etc.)
• qc_fraction: Quality control score
• view_tier: View classification (canonical, etc.)

Metadata Fields (metadata.jsonl)

Full NeuroMorpho.org metadata per neuron:

• neuron_id: NeuroMorpho.org identifier
• species: Species (mouse, rat, human, etc.)
• brain_region: Brain region(s)
• cell_type: Cell type classification
• archive: Source archive/lab
• morphometrics: Quantitative measurements (soma surface, total length, etc.)

Species Distribution

Species	Count	Percentage
mouse	2,000	27.9%
rat	2,000	27.9%
human	1,036	14.5%
chimpanzee	257	3.6%
giraffe	207	2.9%
Other (26 species)	1,658	23.2%

Curation Process

1. Download: Retrieved 11,708 neurons from NeuroMorpho.org
2. QC Filters: Removed failed downloads, degenerate files, zero-bifurcation neurons, integrity errors
3. Outlier Removal: Excluded morphometric outliers (|z| > 5.0)
4. Species Balancing: Capped over-represented species (mouse, rat) at 2,000 each
5. Final Dataset: 7,158 high-quality neurons

Usage

import json

# Load curated manifest (neuron-level, 7,158 neurons)
with open("data/curated_manifest.jsonl") as f:
    neurons = [json.loads(line) for line in f]

print(f"Loaded {len(neurons)} neurons")
# Example: neurons[0] = {"neuron_id": 84160, "species": "African wild dog", ...}

# Load full manifest (image-level, ~164K views)
with open("data/full_manifest.jsonl") as f:
    images = [json.loads(line) for line in f]

print(f"Loaded {len(images)} image records")
# Example: images[0] = {"neuron_id": "10024_ADLR.CNG", "mask": "..._mask.png", ...}

# Load metadata (full NeuroMorpho.org metadata)
metadata = {}
with open("data/metadata.jsonl") as f:
    for line in f:
        record = json.loads(line)
        nid = record.get("neuron_id") or record.get("neuron_name")
        metadata[str(nid)] = record

Related Models

• broadinstitute/axonet-vae-stage1 - Stage 1 VAE for segmentation
• broadinstitute/axonet-clip-stage2 - Stage 2 CLIP for retrieval

Citation

@misc{axonet2025,
  author = {Hall, Giles},
  title = {AxoNet: Multimodal Neuron Morphology Embeddings via 2D Projections},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/broadinstitute/axonet-neuromorpho-dataset}}
}

NeuroMorpho.org Attribution

This dataset is derived from NeuroMorpho.org. Please cite:

Ascoli GA, Donohue DE, Halavi M (2007) NeuroMorpho.Org: A Central Resource for Neuronal Morphologies. J Neurosci 27:9247-9251.

License

• Dataset: CC-BY-4.0
• Source Data: Subject to NeuroMorpho.org terms of use