Data Formats and Representations

Overview

A connectomics dataset is not one thing — it is a family of representations at different levels of abstraction. Raw images, segmentation volumes, surface meshes, morphological skeletons, and connectivity graphs each capture different aspects of the same underlying biology. Choosing the right representation for a given task is a core technical skill, because each format has characteristic strengths, blind spots, and computational costs.

The representation hierarchy

Raw EM images (voxels)
    ↓ segmentation
Labeled volumes (voxel → segment ID)
    ↓ surface extraction
Meshes (triangulated surfaces)
    ↓ skeletonization
Skeletons (tree graphs with spatial coordinates)
    ↓ synapse assignment
Connectome graph (neurons as nodes, synapses as edges)

Each arrow is an information-reducing transformation. You gain computational efficiency and analytical clarity, but you lose spatial detail. The key question is: what information do you need for your analysis, and what is the cheapest representation that preserves it?

Volumetric data

What it is

The most fundamental representation: a 3D array of voxel intensities (raw images) or voxel labels (segmentation). Every spatial position has a value.

Formats

Key properties

• Chunking: Large volumes are divided into chunks (e.g., 128³ or 256³ voxels). Chunks are the unit of I/O — you load one chunk at a time, not the whole volume. Chunk size affects performance: larger chunks = fewer I/O operations but more wasted bandwidth if you only need a small region.
• Multi-resolution pyramids: Store the same volume at multiple resolutions (full res, 2× downsampled, 4×, 8×…). Enables efficient browsing — you see the overview at low resolution and zoom into high resolution on demand.
• Compression: Typical compression ratios of 2-10× for EM data (depending on algorithm: gzip, lz4, zstd, JPEG for lossy). Segmentation volumes compress much better than raw images (large uniform regions).

When to use volumetric data

• Raw image inspection and quality control
• Running segmentation or synapse detection models (need voxel-level input)
• Proofreading (need to see images + segmentation overlay)
• Any analysis requiring spatial context that meshes or skeletons don’t preserve

Limitations

• Storage: A 1 mm³ volume at 4 nm resolution is ~10^13 voxels, ~10 TB at 8-bit. With segmentation (32-bit or 64-bit labels), double or quadruple that.
• Query efficiency: “Which neurons are within 10 μm of this synapse?” requires scanning voxels unless you also maintain a spatial index.

Surface meshes

What they are

Triangulated surfaces that represent the boundary of each segmented object. Each mesh is a set of vertices (3D points) and faces (triangles connecting vertices).

How they’re generated

Marching cubes algorithm (or variants) applied to the segmentation volume. For each segment, extract the isosurface at the boundary between that segment and its neighbors. The result is a watertight mesh (ideally).

Formats

Key properties

• Level of detail (LOD): Store meshes at multiple simplification levels. Full-resolution meshes for a large neuron can have millions of triangles — impractical for real-time rendering. Decimated meshes (10K-100K triangles) are used for overview visualization.
• Vertex attributes: Meshes can carry per-vertex data (e.g., distance from soma, local curvature, synapse density) for visualization and analysis.

When to use meshes

• 3D visualization of neuron morphology
• Surface area and volume measurements
• Spine detection (local curvature analysis on dendritic surfaces)
• Spatial proximity analysis between neurons
• Proofreading — 3D mesh view reveals impossible morphology (merge errors) that is hard to see in 2D slices

Limitations

• Lose internal structure (organelle distributions, cytoplasmic features)
• Mesh topology errors (self-intersections, holes) can arise from noisy segmentation boundaries
• Large storage for complex neurons (a single pyramidal cell mesh can be >100 MB at full resolution)

Skeletons

What they are

Tree-graph representations of neuron morphology. Each skeleton is a set of nodes (3D coordinates along the neurite centerline) connected by edges (parent-child relationships). The root is typically the soma, and branches represent dendrites and axons.

How they’re generated

• From volumes: Thinning/skeletonization algorithms (e.g., TEASAR — Sato et al. 2000) reduce the volumetric segment to its medial axis.
• From meshes: Contract the mesh surface to extract the centerline.
• Manual tracing: Historically, skeletons were traced manually in tools like CATMAID.

Formats

Key properties

• Compact: A neuron that occupies millions of voxels in volumetric form is represented by thousands of skeleton nodes (~KB vs GB).
• Topologically explicit: Branch points, terminal points, and path lengths are directly readable.
• Morphometric analysis: Cable length, branch order, Strahler number, bifurcation angles, tortuosity — all computed directly from skeletons.
• Radius information: SWC format includes radius at each node, preserving approximate process caliber.

When to use skeletons

• Morphological analysis (total cable length, branch complexity, Sholl analysis)
• Cell-type classification based on morphology
• Path-length measurements between synapses
• Input to biophysical simulation (NEURON, Brian)
• Efficient error detection (skeleton shows impossible topology)

Limitations

• Lose surface geometry: Spine morphology, surface area, local curvature not captured
• Lose volume information: Can’t compute volume-based measurements
• Skeletonization errors: Thin processes may be skipped, branch points mislocated, spurious branches created from noisy segmentation
• Radius approximation: SWC radius is a single value per node (circular cross-section assumption), which doesn’t capture irregular shapes

Connectome graphs

What they are

The highest-level representation: neurons as nodes, synaptic connections as edges. This is the “connectome” — the wiring diagram.

How they’re constructed

1. Each segmented neuron = one node
2. Each detected synapse → identify pre-synaptic and post-synaptic segments → create directed edge from pre to post
3. Aggregate: multiple synapses between the same pair → edge weight = synapse count (or sum of cleft areas)

Formats

Node attributes

• Cell type (morphological or transcriptomic classification)
• Soma position (x, y, z)
• Morphological features (cable length, spine density, arbor volume)
• Functional properties (tuning curves from calcium imaging, if available)

Edge attributes

• Synapse count
• Total cleft area or PSD area
• Synapse type (excitatory/inhibitory)
• Spatial locations of individual synapses
• Confidence score

When to use graphs

• Connectivity analysis (degree distributions, clustering, motifs)
• Circuit identification (find all neurons in a pathway)
• Comparison across datasets or conditions
• Input to network models (spiking simulations, dynamical systems)

Limitations

• Lose all spatial information (unless node/edge positions are stored as attributes)
• Lose morphological detail — a graph edge between two neurons doesn’t tell you whether the synapse is on a proximal dendrite or a distal spine
• Thresholding dependence — decisions about minimum synapse count for an “edge” dramatically affect graph structure
• Error amplification — segmentation and synapse detection errors both corrupt the graph

Worked example: choosing a representation

Question: “Do inhibitory interneurons preferentially target the perisomatic region of pyramidal cells in layer 2/3?”

Analysis needs:

1. Identify inhibitory and excitatory neurons → need cell-type labels (graph node attributes)
2. Find synapses between inhibitory → pyramidal pairs → need connectome graph edges
3. Determine synapse location on the pyramidal cell (perisomatic vs distal dendrite) → need synapse spatial coordinates mapped onto the pyramidal cell morphology

Representation choice: This question requires the connectome graph (for connectivity) plus skeletons (for distance-from-soma measurement at each synapse location). Neither the graph alone (no spatial synapse info) nor the volume alone (too expensive for the network-level query) would suffice.

Skeletonization: the method changes the answer

Every skeleton is an approximation of a shape, and different approximations disagree. This matters more than it sounds, because morphology statistics are computed from the skeleton, not from the neuron — so path length, branch count and tortuosity are properties of a method as much as of a cell.

The practical consequence: a morphology statistic is only comparable within one skeletonization method, at one parameter setting. If you compare branch counts between two datasets skeletonized differently, you are partly measuring the difference between the algorithms. State the method and its parameters alongside any morphometric result, exactly as you would state a proofreading level alongside a connectivity result.

Spines are where the methods diverge most, and they are also where the biological question usually is. If spines matter to your analysis, check what your skeletonizer does to them on a handful of cells you have looked at by eye before trusting it on ten thousand you have not.

What each conversion loses

The representations are not interchangeable, and the pipeline between them is one-way. Knowing which direction you are traveling tells you what you can still ask.

The last row is the important one. By the time you are working with an adjacency matrix, every judgment made upstream — segmentation, agglomeration, synapse detection, proofreading — has been baked in and is invisible. That is why provenance metadata is not bureaucracy: it is the only remaining link between a number and the evidence for it.

A round trip is not lossless. Meshing a segmentation and re-voxelizing the mesh does not return the original labels; the surface has been resampled and smoothed. If you need the voxels, keep the voxels.

Storage, in proportion

Format choices look abstract until they meet a storage budget. For a mm³-scale volume, the orders of magnitude are roughly as Unit 04 §2 sets them out: raw and aligned imagery in the petabytes, segmentation labels in the hundreds of terabytes with label-aware compression, meshes in the low terabytes, and skeletons in the tens of gigabytes.

The ratio worth remembering is the last one: skeletons are around five orders of magnitude smaller than the imagery they came from. That is why skeletons get archived and shared while imagery gets left in the bucket it was reconstructed in, and why a great deal of comparative morphology work is possible for anyone with a laptop.

Chunk size is the other lever, and there is no correct value. Small chunks make random access cheap and multiply the number of objects, which object stores handle badly; large chunks reverse both. Pick from your access pattern — a viewer paging through xy planes and an analysis streaming along z want different shapes — and expect to shard the result.

Common misconceptions

References

• Dorkenwald S et al. (2024) “CAVE: Connectome Annotation Versioning Engine.” Nature Methods. doi:10.1038/s41592-024-02426-z.
• Sato M et al. (2000) “TEASAR: Tree-structure extraction algorithm for accurate and robust skeletons.” Proc. Pacific Conference on Computer Graphics and Applications.
• Rubinov M, Sporns O (2010) “Complex network measures of brain connectivity: Uses and interpretations.” NeuroImage 52(3):1059-1069.
• Scheffer LK et al. (2020) “A connectome and analysis of the adult Drosophila central brain.” eLife 9:e57443.