Connectome Quality

What Connectome Quality Means

A reconstructed connectome is a claim: that these objects are neurons, that these contacts are synapses, and that the wiring diagram derived from them can carry scientific weight. Quality control is the practice of measuring how far that claim can be trusted, and no single number does it. Each standard metric measures one thing and is blind to another — the choosing is covered in Unit 08 and the mathematics in the Metrics and QA reference.

Variation of information (VI) totals the disagreement between two segmentations and decomposes into a split component and a merge component. It misleads whenever it is reported as a single number: the split component usually dominates, so a total VI can improve while merge errors — the expensive kind — get worse. Report the two components separately, always.

Expected run length (ERL) is the mean error-free path length along a neuron’s skeleton: how far can you trace before hitting an error? It fits tracing-oriented questions, and misleads on merges, which it does not penalize unless explicitly made to.

Synapse precision and recall score detected synapses against ground truth. They assume the segmentation underneath is correct — a synapse assigned to a merged object still scores as a hit — so they can look excellent on a volume whose wiring diagram is wrong.

Completeness reports what fraction of a neuron was reconstructed, and says nothing about whether what is there is correct.

Behind all four sits the field’s central asymmetry: splits are visible and bounded, while merges are invisible and unbounded, which is why pipelines are deliberately tuned to over-segment. And behind the metrics sits the question they exist to serve — not “what is our VI?” but “how much would our result change under correction of the remaining errors?” Unit 08’s answer is to exhaustively proofread a small random sample of analysis cells (about 20 is often informative) and report how the endpoint shifts.

These methods are exercised on real projects: MICrONS provides densely reconstructed EM volumes with functional data that serve as a testbed for quality assessment, and CIRCUIT (Connectome Integrity and Reliability through Quantitative and Iterative Training), developed by William Gray-Roncal and collaborators, packages evaluation tools and metrics — topology, morphology, and synapse-based F1 score — for scalable use.

How Humans and Machines Divide the Work

Automated segmentation produces the reconstruction; automated detectors then propose where it is wrong — endpoint detectors flag likely splits (a neurite that stops in mid-neuropil), implausible-morphology detectors flag likely merges (organelle and shape combinations that cannot coexist in one process). The output is a ranked queue of candidates, not a verdict.

Humans adjudicate that queue. The division holds because the two error types demand different strengths: splits are findable by rule, but recognizing a merge requires the biological judgment that a “perfectly ordinary looking” object is in fact two cells — which is why Unit 08 assigns merge classes to human review. Newcomers enter this workflow through small, atomic proofreading tasks — validating boundaries in bounded regions, with every correction carrying an evidence chain — and that is a deliberate training design as much as a labor one: the proofreading side quest builds the skill on real public volumes, and structured task designs with built-in quality checks let larger groups contribute without diluting the standard.

</section>

<div class="cards-grid">
    <div class="card">
        <h2>Learn by Doing</h2>
        <p>The hands-on route is the <a href="/side-quests/proofreading/">proofreading side quest</a>: worked scenarios, real public volumes, and an artifact a lab can read. The <a href="/notebooks/connectome-quality/">notebooks page</a> holds the reference-code steps for computing the metrics on this page.</p>
    </div>

    <div class="card">
        <h2>Where to go from here</h2>
        <ul>
            <li><a href="/technical-training/08-segmentation-and-proofreading/">Unit 08</a> — the metrics on this page taught in full, with a graded lab</li>
            <li><a href="/side-quests/proofreading/">The proofreading side quest</a> — the hands-on route, on real public volumes</li>
            <li><a href="/ask-an-expert/">Ask an Expert</a> — the site's question route</li>
        </ul>
    </div>
</div>

</div>