Neuroglancer integration overview

Neuroglancer is a WebGL-based viewer for petascale volumetric data, widely used in connectomics, brain imaging, and synchrotron science. It natively understands several data formats (OME-Zarr image volumes, Neuroglancer precomputed, N5), but does not have native support for the Zarr Vector Format.

zv-ngtools bridges this gap. It is a fork of ngtools — a collection of Neuroglancer utilities — extended with a zarr_vectors layer type that translates ZVF stores into Neuroglancer layers on the fly.

Why Neuroglancer cannot read ZVF natively

Neuroglancer’s data source plugins expect data in specific binary formats (precomputed, N5, OME-Zarr). ZVF stores vertex data as spatially chunked Zarr arrays with a fragment index — a structure Neuroglancer does not understand without a mediating translation layer.

zv-ngtools provides that layer: it runs a local HTTP file server that intercepts Neuroglancer’s chunk requests, reads the requested spatial region from a ZVF store using zarr-vectors-py, and returns the data in a format Neuroglancer expects.

Two integration paths

Path A — Local serving via LocalNeuroglancer

A Python process (zv-ngtools) runs a local Tornado HTTP server alongside a Neuroglancer browser tab. The server:

  1. Receives chunk requests from Neuroglancer (HTTP range requests).

  2. Uses zarr-vectors-py to read the requested fragment slices.

  3. Translates the ZVF data to the Neuroglancer layer protocol.

  4. Returns the response to the browser.

LOD is driven by the ZVF resolution pyramid: as the user zooms out in Neuroglancer, the server switches to coarser levels automatically, using the bin_shape at each level to select the appropriate resolution.

Python process                         Browser
┌───────────────────────────────┐     ┌─────────────────────────────┐
│  LocalNeuroglancer             │     │  Neuroglancer (WebGL)        │
│  ├─ Tornado HTTP server        │◄────│  chunk requests              │
│  │   (port 9123)               │────►│  rendered layers             │
│  └─ zarr-vectors reader        │     │                              │
│      ├─ scan.zarrvectors       │     └─────────────────────────────┘
│      └─ tracts.zarrvectors     │
└───────────────────────────────┘

This path is best for local analysis and collaborative review sessions where the data lives on local disk or a mounted network share.

Path B — Precomputed export for static hosting

Convert a ZVF store to the Neuroglancer precomputed format and upload to a public HTTP server (S3, GCS, nginx). Neuroglancer fetches the data directly without any intermediary Python process.

This path is best for sharing published datasets with collaborators who will not install zv-ngtools, or for embedding Neuroglancer links in publications and websites.

See Precomputed export for the conversion workflow.

What zv-ngtools adds over upstream ngtools

zv-ngtools is a fork of ngtools (neuroscales/ngtools). All upstream ngtools functionality is preserved; the fork adds:

Feature

Upstream ngtools

zv-ngtools

Load OME-Zarr image volumes

✓

✓

Load TRK / TCK tractography

✓

✓

Load NIfTI, MGH, TIFF

✓

✓

Load .zarrvectors stores

✗

✓

ZVF-aware LOD selection

✗

✓

zarr_vectors layer type

✗

✓

Precomputed export from ZVF

✗

✓

Point cloud layer rendering

✗

✓

Streamline layer rendering

✗ (only TRK/TCK)

✓ (from ZVF)

The zarr:// URL scheme (for OME-Zarr stores) works identically in both forks. Only the zarr_vectors:// scheme is new.

Layer type mapping

ZVF geometry type

Neuroglancer layer type

Notes

point_cloud

annotation (point)

Rendered as 3-D points; size and colour from attributes

line

annotation (line)

Rendered as line segments

polyline / streamline

annotation (line)

Rendered as connected line sequences

graph / skeleton

segmentation (mesh-less) or custom skeleton layer

SWC-style rendering

mesh

segmentation (mesh) or annotation (surface)

Draco-compressed meshes supported

For types without a perfect Neuroglancer native equivalent (e.g. general graphs), zv-ngtools uses the closest Neuroglancer layer type and adds a server-side translation step.

Coordinate system alignment

ZVF stores and Neuroglancer image volumes can be displayed together when they share a coordinate system. Neuroglancer uses a global coordinate space for all layers; each layer specifies its own coordinate transform.

zv-ngtools reads the coordinate_system and multiscales.axes fields from the ZVF store’s .zattrs to set the Neuroglancer layer transform. Stores declared in RAS (Right-Anterior-Superior) space are automatically aligned with Neuroglancer’s default RAS coordinate system.

For stores in voxel space, pass an explicit affine when loading:

viewer.add(
    "scan.zarrvectors",
    transform=np.array([
        [0.004, 0,     0,     0],     # 4 µm per voxel
        [0,     0.004, 0,     0],
        [0,     0,     0.025, 0],     # 25 µm z-step
        [0,     0,     0,     1],
    ]),
)

Quick-start checklist

Before working through the detailed tutorials:

  • pip install git+https://github.com/BRIDGE-Neuroscience/zv-ngtools.git

  • A web browser that supports WebGL 2 (Chrome, Firefox, Edge)

  • A .zarrvectors store with at least a base level (level 0)

Optionally, for best performance:

  • A multi-level pyramid (run build_pyramid first)

  • Consolidated metadata (zarr.consolidate_metadata)