Understanding Edit Scope: Segments, Supervoxels, and Root Identifiers

Welcome. This module shifts from finding suspicious reconstructions to changing them safely and keeping those changes interpretable later. The key idea is deceptively simple: a visible “neuron segment” is not a single indivisible object. It is a current grouping of smaller, persistent units, and an edit changes that grouping.

By the end of this lesson, you should be able to explain what FlyWire means by a supervoxel, a segment, and a root ID; predict what happens to identifiers after a merge or split; and judge the real scope of an edit before committing it.

---

The three levels: image, supervoxel, current segment

Electron microscopy supplies the underlying image volume: many adjacent 2D sections that together show membranes, cytoplasm, organelles, and processes. Segmentation software assigns regions in that image volume to candidate objects. FlyWire does not treat the resulting reconstruction as one monolithic mesh that is overwritten whenever someone proofreads it.

Instead, its segmentation is built from supervoxels.

Here is the practical hierarchy:

Level	What it is	Does it change when someone proofreads?
EM voxels	The underlying image samples, containing visual evidence	No
Supervoxel	An atomic region created by automated segmentation	No; it is immutable
Root ID	An identifier for one particular grouping of supervoxels	The old ID is superseded by a new one after an edit
Displayed segment	What you typically experience as the selected reconstructed object in the interface	Its current membership may change after edits

The word segment is useful but context-dependent. In the interface, it commonly means the visible reconstructed object you clicked. Technically, that object is represented by a root ID, which denotes a collection of supervoxels. When precision matters—especially in notes, reviews, or code—say which one you mean:

• “The selected root includes an incorrectly attached branch.”
• “The candidate branch is composed of different supervoxels.”
• “This displayed segment is likely a false merge.”

That wording prevents a common beginner error: assuming that clicking one small-looking piece means that only a small-looking piece is affected.

The hierarchy shown in Hierarchical segmentation and correction matters because it separates two ideas:

1. The pieces are stable. A supervoxel remains a known atomic unit.
2. The grouping is revisable. FlyWire can change which supervoxels belong together in a current reconstruction.

This is why proofreading can refine a huge connectome without redrawing every cell manually.

---

A root ID is a versioned reconstruction, not a permanent neuron name

A root ID identifies a particular state of a reconstruction. It should not be treated as the eternal identity of a biological neuron.

Suppose a current reconstruction has root ID $R$ and contains a set of supervoxels:

$$R=\{s_1,s_2,\dots ,s_n\}$$

A nearby disconnected fragment has root ID $Q$:

$$Q=\{t_1,t_2,\dots ,t_m\}$$

If evidence shows that they are two pieces of the same neurite, a merge produces a new root ID, say $R^{\prime }$, representing their combined membership:

$$R^{\prime }=\{s_1,s_2,\dots ,s_n,t_1,t_2,\dots ,t_m\}$$

The important operational consequence is that $R$ and $Q$ are no longer the current roots for those objects. The underlying supervoxels still exist, but the prior root IDs have been superseded.

A split has the symmetric identifier consequence. If one current root mistakenly contains material from two neurons, a successful cut separates its supervoxels into two new collections:

$$R_A=\{s_1,\dots ,s_k\}$$ $$R_B=\{s_{k+1},\dots ,s_n\}$$

The original root $R$ is then no longer the current root for either resulting object.

Why immutability is useful

At first, “every edit makes a new ID” can seem inconvenient. In fact, it is a valuable design choice.

If the same root ID silently changed its membership over time, it would be difficult to determine what a previous analysis meant. A connectivity query, annotation table, or notebook could appear reproducible while actually referring to a different reconstruction later.

With versioned roots, you can make a defensible statement such as:

At a specified materialization version, this analysis used the current descendant of a reviewed reconstruction.

A materialization is a snapshot of segmentation and associated tables at a particular point in time. For now, retain the principle: a bare root ID without its data version may be insufficient for reproducible analysis. Module 3 will turn that principle into a notebook workflow.

---

Edit type determines scope; visual size does not

The FlyWire 101 guide describes the user-facing idea: the AI has already fused many supervoxels into larger segments, and proofreading corrects errors in those fusions.

The two edit categories have opposite biological intent:

Edit	Reconstruction error being repaired	Membership consequence	Root-ID consequence
Merge	False split: one biological neuron has been separated into disconnected reconstructed pieces	Two root-level collections become one current collection	Two old roots are superseded by one new root
Split	False merge: material from different neurons is grouped together	One root-level collection is divided into separate current collections	One old root is superseded by two or more new roots

The scope of an edit has two dimensions.

1. Local operation scope

The interaction is local: you place merge points or split-point labels in a limited region of the EM data. This local evidence must be strong. A membrane boundary, branch continuity, and agreement across adjacent sections matter more than a plausible-looking 3D silhouette.

For a merge, the immediate question is: Are these two nearby supervoxel groups truly parts of one neurite?

For a split, it is: Do the selected points correctly distinguish the two biological objects at the erroneous connection?

2. Global reconstruction scope

The consequences may be much larger than the local interaction.

A merge at the tip of a short branch can unite two otherwise large reconstructions. A split made at a tiny contact point can detach an entire arbor, axonal continuation, or unrelated cell body from the current root. The apparent size of the error in one 2D section is therefore not a reliable measure of the size of the resulting change.

A useful mental model is graph editing:

• Supervoxels are the stable units in the graph.
• The current root is the connected agglomeration that FlyWire presents as one reconstruction.
• A merge asserts that two previously separate agglomerations belong together.
• A split removes an erroneous connection so that distinct components can be represented separately.

You do not need to understand FlyWire’s internal graph implementation to proofread responsibly. You do need to recognize that a click is an assertion about membership, not merely a cosmetic change to a mesh.

---

Worked cases: predict the new roots before editing

Case A: repairing a false split

You inspect a neurite that ends abruptly at location $L$. Across several image sections, a nearby segment continues with compatible caliber, texture, and direction. Suppose the selected root contains the main arbor, while the nearby fragment is another root.

A justified merge claims that the two root-level objects are one biological reconstruction. Its expected outcome is:

• one current root containing both collections of supervoxels;
• both pre-edit root IDs becoming historic;
• the merged morphology requiring a fresh 2D and 3D inspection.

The critical caution is that “nearby” is not enough. In dense neuropil, adjacent neurites can run parallel or cross without being connected. This is why the next lesson focuses on verifying evidence across sections and in 3D before an edit.

Case B: repairing a false merge

You find a root that mostly resembles one neuron but contains an attached branch with incompatible morphology. In the EM volume, the connection appears to cross a membrane boundary, indicating that the reconstruction bridges two cells.

A split does not merely erase the visible junction. It changes the root grouping so that the involved supervoxels resolve into separate current roots. You must anticipate both results:

• the intended neuron should retain its coherent morphology;
• the detached object should be a plausible, inspectable reconstruction rather than an accidental collection of fragments.

In a complex false merge, the local attachment may have several contacts or an intertwined geometry. The FlyWire 101 guide introduces simple and multi-cut splitting tools, but tool choice must come after evidence review. A sophisticated tool does not compensate for uncertain biological interpretation.

Case C: a plausible-looking but unsafe merge

Imagine a small fragment visually touches the end of a target dendrite in one section. A single image can be misleading: a membrane may be faint, oblique, or obscured by staining variation. If the adjacent sections show the fragment passing behind the dendrite rather than continuing into it, merging would create a false merge.

The correct outcome here is no edit yet. Record the location and uncertainty, then inspect more evidence or seek review. Restraint is a valid proofreading decision because each edit affects both morphology and the identifier history used by later analysts.

---

A pre-edit scope statement

Before you eventually press an edit control, make a brief private statement of the change you expect. This is a compact way to catch misunderstandings early.

For a proposed merge:

I expect current roots $R$ and $Q$ to be replaced by one new root. The new root should contain the target arbor and the candidate continuation, with no additional unrelated morphology.

For a proposed split:

I expect current root $R$ to be replaced by separate roots. One should retain the target cell body and main arbor; the other should contain the incorrectly attached branch.

Then check four things:

1. Which objects are involved? Identify the currently selected root and the other root or branch at the candidate site.
2. What is the biological claim? State whether this is one neuron wrongly separated or two neurons wrongly combined.
3. What is the expected scope? Consider whether a large branch, distant arbor, or cell body could be part of the affected component.
4. What identifiers and evidence need recording? Preserve the relevant root IDs, location, and the materialization or dataset context when available.

This discipline resembles version-aware work in software: an operation may be triggered in one small file or function, but its real impact can propagate through shared dependencies. Here, the dependencies are the supervoxel groupings that make up a root.

Do not infer that an old root ID is useless. Historic roots can still be useful for understanding what existed before an edit. But do not mix a historic root with current roots as though all were simultaneous, current reconstructions. That can lead to invalid metadata or connectivity interpretations.

---

Key takeaways

• Supervoxels are FlyWire’s atomic, immutable segmentation units.
• A visible segment is typically a current grouping of many supervoxels, represented by a root ID.
• A merge repairs a false split by joining root-level collections; a split repairs a false merge by separating one collection.
• The interaction may be local, but the morphological and identifier consequences can be global.
• Root IDs are immutable: after an edit, the old root or roots are superseded by newly created root IDs.
• Record data-version context when work needs to support later querying or analysis.

Next, you will learn how to verify a proposed edit across adjacent EM sections and in the 3D reconstruction, so that the edit-scope reasoning from this lesson rests on visual evidence rather than intuition alone.