H01 electron microscopy with object segmentation and original 2 µm scale bar

Connectomics Ethics and Governance

A graduate lecture from the NeuroTrailblazers reference layer

NeuroTrailblazers · neurotrailblazers.org

Part A — Human tissue: consent and de-identification
Part B — Licences: what a reuser must do
Part C — Dual use and credit for proofreading

Human cortex · H01
Object segmentation over electron microscopy

H01 release · Lichtman Lab / Harvard & Connectomics at Google
Image: CC BY 4.0 · Shapson-Coe et al. (2024) · doi:10.1126/science.adk4858
Lecture: CC BY-SA 4.0 · neurotrailblazers.org

Start with the honest assessment:

connectomics does not currently have a human-subjects problem.

Its flagship human dataset is discarded surgical tissue that was going to be removed regardless, reviewed under the same frameworks that govern any human-tissue study. Nothing in this field asks a person to undergo a procedure for research, and nothing published so far identifies anyone.

If you came expecting a controversy, there isn't one to report. So what is this hour for?

What is live now, and what is prospective

Live now — practical, not philosophical

Licence obligations (Part B). Three major portals redistribute under incompatible terms, and a reuser who ignores them is simply in breach.

Credit for proofreading labour (Part C). Tens of thousands of hours of human correction sit behind every published connectome, and how that work is credited decides real careers. The one topic here about people affected today.

Prospective — worth discussing because it is not yet urgent

  • Consent framings written for a cubic millimetre of discarded tissue, not for whole-brain human volumes or machine-learning corpora
  • De-identification: a non-issue at today's volumes, an open question at tomorrow's
  • Neural-data regulation: UNESCO adopted the first global standard in November 2025
  • Dual use: weak for a field with no pathogens, and should be stated at its real strength

Treat the prospective material as a seminar, not a compliance checklist. The right time to reason about it is before a project needs the answer.

Learning objectives

By the end of this lecture you will be able to:

1 — Locate the provenance and ethics statement of a human-tissue connectomics dataset before citing it.

2 — Explain where the residual re-identification risk in a nanoscale EM volume actually lives.

3 — Determine what a dataset's licence obliges a reuser to do, and whether a planned use is permitted.

4 — Distinguish governed, arguable and actual dual-use concerns, at their real strength.

5 — Compare credit models for proofreading labour, and state what each costs the contributor.

Objective 3 is the material people skip and then get wrong. In at least one flagship case, the paper and the data carry different licences.

Roadmap

Part A — Human tissue (prospective) Where H01 came from; what its paper does and does not say; oversight regimes compared; what an EM volume can and cannot reveal; the UNESCO Recommendation.

Part B — Licences (live now) Five resources, five licence positions; the FlyWire row; funder and repository layers; a decision table.

Part C — Dual use and credit (one prospective, one live) Three dual-use concerns at their real strength; the scale of proofreading labour; four credit models in use; what to do about it.

Bring to each part: a connectomics dataset you have used or plan to use. By the end you should know where its ethics statement is, what its licence lets you do, and how its proofreaders were credited.

Part A

  • Where H01 came from, and what its paper says
  • Three oversight regimes, compared
  • What an EM volume can and cannot reveal
Slides 6–14

Where H01 came from

The site's human reference volume has a specific and documentable provenance

Shapson-Coe et al. (2024) describe the sample as a "rapidly preserved, 170-µm-thick slab of human cortex from the anterior part of the middle temporal gyrus of a 45-year-old female", of "just over 1 mm³", which "was removed to gain access to an epileptic focus in the underlying hippocampus".

Step Figure
Sections 5,019, mean thickness 33.9 nm
Imaging Multibeam SEM at 4 × 4 nm
Dataset About 1.4 petabytes
Content Roughly 57,000 cells and 150 million synapses

Shapson-Coe et al. 2024, Science 384, eadk4858 (10.1126/science.adk4858).

A surgical by-product

The strongest position human connectomics can currently occupy

In the authors' framing, human specimens become available "from neurosurgical interventions for neurological conditions in which pieces of the cortex are discarded because they obstruct access to a pathological site".

No tissue was removed for research that would not otherwise have been removed.

That is why H01 — rather than any purpose-collected human sample — is the field's reference volume.

The ethical structure follows directly from the provenance description on the previous slide.

What the paper does and does not say

The main article carries no consent or IRB statement.

A full-text search of the published Science article for "consent", "IRB", "ethics" and "Institutional Review" returns nothing. The data-availability statement points at the release landing page, and the licence line covers the article, not the data.

The consent and approval details live in the supplementary Materials and Methods.

The practice to adopt.

When you reuse human tissue data, find the ethics statement before you cite the dataset, and say where you found it.

A citation to a landing page is not a citation to an ethics review.

The paper is explicit about the sample's pathology

Human surgical samples "originate in individuals with pathologies of the nervous system such as epilepsy, tumors, or neurodegenerative diseases". In this case, "we cannot exclude the possibility that long-term epilepsy, or its pharmacological treatment, had subtle effects on the nanometer-scale structure of the tissue".

This is simultaneously a scientific caveat and an ethical one.

Every claim of the form "the human brain does X, as shown by H01" is really a claim about tissue from one adult woman with drug-resistant epilepsy.

Part C returns to why that matters beyond the lab.

Three oversight regimes, compared

Source Oversight regime What consent covers
Human surgical tissue (H01) Institutional review of human-subjects research; tissue is discarded surgical material Research use of the specimen. Whether "research use" was understood to include indefinite open redistribution and downstream machine-learning use is the live question, not a settled one
Vertebrate animal tissue (MICrONS, MouseConnects) Institutional animal care and use committee approval Protocol-level: species, numbers, procedures, endpoints
Invertebrate tissue (FlyWire, hemibrain) No animal-welfare committee requirement in the United States for Drosophila Not applicable

The asymmetry is the point. The dataset with the fewest formal gates — Drosophila — is the one the field proofreads most openly and credits most carefully. The dataset with the most gates is the one whose ethics statement is hardest to find.

De-identification: what an EM volume can and cannot reveal

Structural MRI has a well-known re-identification route. A head volume can be rendered as a face, which is why defacing is a standard preprocessing requirement.

Nanoscale EM has no equivalent route. A 1 mm³ block of cortex imaged at 4 nm contains no face, no skull shape, no name, no date, and none of the direct identifiers a health-privacy framework enumerates.

The image data carries no genotype: what is released is heavy-metal-stained, resin-embedded tissue rendered as grayscale voxels.

The residual risk is contextual, not pictorial — and it lives in the metadata, not the images:

  • Population size. "A 45-year-old woman who had an anterior temporal resection for drug-resistant epilepsy, at a named centre, in a narrow window of years" is a small set of people.
  • Linkage. Dates, site and clinical detail combine with other records in ways no single field does alone.
  • Pathology. Features in the tissue reflect the donor's condition, and the condition is part of why the tissue exists.

Two consequences for practice, and where governance is moving

1 — De-identification in EM connectomics is metadata discipline, not image modification. There is nothing to blur.

2 — The standard reference transfers only partly. Betzel & Bhatt (2021) on responsible reuse of open human connectomics data was written about neuroimaging-scale datasets. Its reporting and analytic-flexibility guidance transfers cleanly. Its privacy guidance was designed for a modality with a face in it.

UNESCO Recommendation on the Ethics of Neurotechnology — adopted by the General Conference on 12 November 2025, the first global standard in the area.

  • establishes a framework for "neural data"
  • holds that such data is uniquely sensitive and requires strict safeguards against misuse
  • is non-binding on member states

Whether a volume EM dataset falls inside that definition of neural data is not settled, and this lecture does not claim it does.

Check yourself

Why is "we removed the patient's name" not the whole of de-identification for H01 — and what is the residual risk?

Because the volume never contained a name — or a face, or any direct identifier. There is nothing in the image to remove.

The residual risk is contextual. The clinical description — age, sex, procedure, site, approximate date — picks out a small population, and it lives in metadata and prose, not in voxels.

Follow-up: which details of H01's provenance description — age, sex, procedure, site — would you leave out of a public dataset record, and what would a reuser lose if you did?

Part B

Licences: what a reuser is actually obliged to do

  • Five resources, five licence positions
  • The row that catches people
  • Funder and repository layers, and a decision table
Slides 15–20

The section people skip and then get wrong

Resource Licence What you must do Commercial use
MICrONS (microns-explorer.org) Creative Commons Attribution 4.0 International Retain creator identification and copyright notice, indicate modifications, include the licence text or a link; impose no further restrictions downstream. Follow the site's citation policy — MICrONS Consortium et al. (2025), Nature 640: 435–47 Permitted
FlyWire public release (v783, an October 2023 snapshot) Creative Commons Attribution-NonCommercial 4.0 Attribute, and cite the papers named in FlyWire's citation guide. Pre-publication data carries separate community-principles obligations Not permitted
FlyWire's flagship paper (Dorkenwald et al., 2024, Nature) Creative Commons Attribution 4.0 Attribute Permitted
Hemibrain / neuPrint Creative Commons Attribution (version not confirmed from the primary source — check before redistributing) Attribute; cite Scheffer et al. (2020) Presumed permitted, unverified
H01 No licence statement found on the release landing page; the Science article is under the AAAS journal licence, © the authors. Data availability points at the public Google Cloud bucket, open with no registration Cite Shapson-Coe et al. (2024) and check terms before redistributing Unclear — verify

In at least one case the paper and the data carry different licences. Two rows are marked unverified rather than guessed.

The FlyWire row is the one that catches people

Reusing a figure from the Nature paper and reusing the connectome are governed by different instruments.

Licence
The article CC BY 4.0
The data release CC BY-NC 4.0

A commercial product built on FlyWire connectivity is a licence problem.

A commercial product built on MICrONS connectivity is not.

Your lab spins out a company that sells a connectome-constrained model trained on FlyWire. Which licence governs, and what does it say?

Two more layers above the per-dataset licence

Funder obligations bind the producer, not the reuser.

The NIH Data Management and Sharing Policy (NOT-OD-21-013) took effect on 25 January 2023. Every NIH application that will generate scientific data must include a data management and sharing plan, and comply with it.

That is why CONNECTS-scale projects release data at all — but it grants a reuser nothing beyond what the dataset's own licence grants.

Repository policies span a spectrum.

Jwa & Poldrack (2022) surveyed data-sharing policies across neuroimaging repositories. Access control and restrictions on secondary use vary widely — from fully open, download-and-reanalyse without constraint, to controlled access requiring verified credentials and limiting permitted secondary analyses.

The practical lesson: "the data is public" is not a licence statement.

A decision table

What to check, by what you want to do

You want to… Check
Put a published figure in your paper or slides The article licence, not the data licence
Redistribute a derived table (e.g. an edge list) The data licence plus its attribution requirements
Train a model on the imagery The data licence; a NonCommercial clause reaches the model if the model is commercial
Build a product or a paid service Whether any NC clause applies — FlyWire yes, MICrONS no
Publish a reanalysis The portal's citation policy: cite the papers it names, not only the URL
Quote a number in a lecture Nothing — but pin the release version you used

Note the last row. Licensing is not the only obligation that travels with a number. Pinning the release version is the reproducibility half of governance.

Check yourself

You want to build a paid teaching product around a connectome. Which of MICrONS and FlyWire can you use, and why?

MICrONS — yes. It is CC BY 4.0, which permits commercial use provided you attribute, mark modifications, and pass the licence on.

FlyWire's data — no. The data release is CC BY-NC 4.0, so a paid product is outside the licence — even though FlyWire's Nature paper is CC BY and its figures are reusable.

And H01 or hemibrain? Hemibrain is presumed permitted but its CC BY version is unverified; H01's data licence is unclear. For both, the only defensible answer is "verify before redistributing".

Part C

Dual use and credit for proofreading

  • Three concerns, at their real strength
  • The scale of proofreading labour
  • Four credit models, and what to do about it
Slides 21–27

Dual use, honestly

Connectomics has none of the hazards the classic dual-use frameworks were built for

There is no pathogen, no agent, no enhancement of transmissibility. Saying otherwise inflates the field's risk profile and wastes the reader's attention. Three concerns are real, and only one of them is currently governed.

Status Concern At its real strength
Governed Neural-data misuse The UNESCO Recommendation situates neurotechnology within a human-rights framework — dignity, freedom of thought, mental privacy, autonomy — and calls on states to prevent applications that facilitate coercive control, unlawful surveillance or manipulation. Not binding, and it primarily targets devices that read or write brain activity in living people, not post-mortem or surgical structural datasets
Arguable, undocumented Method transfer Segmentation, tracking and instance-association methods developed for connectomics are general computer vision, which has surveillance applications. An argument from capability, not from any documented case — recorded as an argument, not a finding
Actual Over-claiming from a single sample The one a student is likely to commit. Next slide

The actual one: over-claiming from a single sample

"H01 shows that the human brain does X" is a sentence about tissue from one person with epilepsy (Part A).

Extended into difference or disease narratives, a claim of that shape is a scientific integrity failure with social consequences well outside the lab.

The mitigation is not a policy.

It is writing the sample description into the claim, every time.

Weak: "The human cortex shows X."

Defensible: "In ~1 mm³ of middle temporal gyrus resected from one patient with drug-resistant epilepsy, we observe X."

Credit for proofreading labour: the scale

Dorkenwald et al. (2024) estimate that FlyWire's brain reconstruction took around 33 person-years of manual proofreading, distributed across consortium labs, centralised teams at Princeton and Cambridge, and citizen scientists worldwide.
Community annotations shared 133,700
Proofread neurons 139,255
Synapses 54.5 million

That is the only published effort figure of its kind the source page could find. MICrONS and H01 do not publish an equivalent person-year number.

The absence is itself a finding: the largest single labour input to a connectome is routinely unquantified in the paper that reports the connectome.

Dorkenwald et al. 2024, Nature 634, 124–138 (10.1038/s41586-024-07558-y).

Four credit models actually in use

Model Example What a contributor can put on a CV Failure mode
Consortium co-authorship "The FlyWire Consortium" as a co-author of Dorkenwald et al. (2024) Membership of a named group; the individual is not indexed by name in bibliographic databases Hard to claim in a job application or a tenure case; invisible to citation metrics
Collective acknowledgement in the author line Kim et al. (2014), Nature — the author list ends "and the EyeWirers", with the individuals who reconstructed the cells listed in supplementary information A verifiable named appearance, but in supplementary material Not authorship; disappears from every automated record
Per-contribution platform attribution FlyWire Codex shows per-cell credits and a labelling leaderboard A durable, checkable link to specific work Not a publication; depends on the platform continuing to exist
Named individual authorship with a contributions statement Shapson-Coe et al. (2024) names individuals for "proofreading of neurons", "production of ground truth for synapse prediction and excitatory versus inhibitory classification", and each other task Full authorship plus a specific, quotable role Only workable when the contributor count is small

Every model has a failure mode. Which one would you choose for a project with 40 contributors, and what would you tell them before they started?

What to do about it

Three things that can be stated without hedging

1 — CRediT has no term for proofreading. The nearest contributor-role terms are Data curation and Investigation, and neither describes segment-level error correction. If your project uses CRediT, write down which term you are mapping proofreading onto, before results exist.

2 — The threshold has to be written before the work. "How much proofreading earns authorship?" answered after a paper is drafted is answered under pressure. The four models are the menu; pick one and publish the rule.

3 — Effort is measurable, so measure it. FlyWire's 33 person-years exists because someone counted. Edit histories in CAVE-backed datasets make per-contributor effort computable. A project that does not report it has chosen not to.

Check yourself

A colleague says an undergraduate who proofread 800 segments should be "acknowledged, not an author". Which credit model is that, and what does the contributor lose?

The model: collective acknowledgement — the EyeWire model.

What the contributor loses: authorship, indexing under their own name in bibliographic databases, and any citation credit.

What they gain: a named appearance that most automated records will never see.

That may be the right call.

It should be the stated call, made before the work — not after.

What this lecture does not cover

Boundaries stated deliberately, as the source page states them

  • Legal advice and jurisdiction-specific law. GDPR special-category data, US state genetic-privacy statutes and national neurorights legislation are real and consequential; none is analysed here. Nothing here substitutes for your institution's research-compliance office.
  • Publication ethics in general. COPE Core Practices and the ICMJE authorship criteria are the operative documents, applied in the site's peer-review and ethics module.
  • Animal welfare procedure. IACUC protocol design, the 3Rs and species scope are named and not developed.
  • Licence details not confirmed from a primary source. The hemibrain/neuPrint CC BY version and the H01 data licence are marked unverified rather than guessed. Confirm both before redistributing.
  • Repository-by-repository policy counts. Jwa & Poldrack is cited for its spectrum finding only.
  • Consent for future AI training uses. Whether broad consent to "research use" of surgical tissue extends to training foundation models on the resulting images is open. No settled answer exists, and inventing one would be worse than saying so.
  • Community conduct and inclusion — codes of conduct, harassment policy, mentoring obligations.

Where to go next

The source page. Ethics and Governance in the NeuroTrailblazers content library — the facts in this lecture, with every document linked.

Practising the judgement. The site's peer-review and scientific-ethics module, with a review-board simulation — where these facts become decisions.

Neighbouring pages. The H01 case study (its pathology caveats); the FlyWire case study (the community whose labour Part C counts); Provenance and versioning (pinning the release you used); Synapse detection (a worked example of bounded claims); Datasets (access routes and per-dataset terms).

The one idea to carry forward.

The live obligations are unglamorous: read the licence that actually governs, find the ethics statement before you cite, write the sample into the claim, and settle credit before the work.

The prospective questions are real — and the time to reason about them is before a project needs the answer.

References and sources

Datasets and their papers. Shapson-Coe et al. 2024 (10.1126/science.adk4858, H01); Dorkenwald et al. 2024, and the FlyWire Consortium (10.1038/s41586-024-07558-y, FlyWire); MICrONS Consortium et al. 2025 (10.1038/s41586-025-08790-w), with terms at microns-explorer.org/terms-and-conditions; Scheffer et al. 2020 (10.7554/eLife.57443, hemibrain); Kim et al. 2014, and the EyeWirers (10.1038/nature13240).

Governance and policy. UNESCO (2025), Recommendation on the Ethics of Neurotechnology, adopted 12 November 2025 (unesco.org); NIH (2020), Final NIH Policy for Data Management and Sharing, NOT-OD-21-013, effective 25 January 2023 (grants.nih.gov); FlyWire citation and credit guidelines (flywire.ai/guidelines).

Responsible reuse. Betzel & Bhatt 2021 (10.1016/j.neuroimage.2021.118579, open human connectomics datasets); Jwa & Poldrack 2022 (10.1002/hbm.25803, the spectrum of data-sharing policies).

Source page. NeuroTrailblazers content library, Ethics and Governance (/content-library/connectomics/ethics-and-governance/). https://neurotrailblazers.org

Use, adapt, and credit

These slides are openly licensed for community use

Licence: CC BY-SA 4.0
Creative Commons Attribution-ShareAlike 4.0 International.
https://creativecommons.org/licenses/by-sa/4.0/

You may teach from these slides anywhere, including commercially; copy and redistribute them in any medium; and re-cut, shorten, translate, restyle, or merge them into your own material — and distribute the result. No permission needed.

Two conditions. Attribution — credit the original, link the licence, and say if you changed anything. ShareAlike — distribute your adapted version under this same licence, so it stays as open as what it came from.

How to credit

NeuroTrailblazers (2026). Connectomics Ethics and Governance (graduate lecture, NeuroTrailblazers reference layer). CC BY-SA 4.0. neurotrailblazers.org/technical-training/slides/

For an adaptation, prefix with "Adapted from" and note what you changed.

Editable source. The Marp markdown is in the repository — the exported PowerPoint renders each slide as an image, so the markdown is the thing to edit. https://github.com/wrgr/neurotrailblazers

Improved something? The project would like to hear about it — open an issue.

Cover image: H01 release, Lichtman Lab / Harvard & Connectomics at Google, CC BY 4.0. Shapson-Coe et al. (2024), doi:10.1126/science.adk4858. The image retains its own licence. Cited papers carry their own licences; citation is not reproduction.

This lecture is built from one page of the NeuroTrailblazers content library, "Ethics and Governance", and every fact on these slides comes from that page and the documents it cites. It holds facts, not decisions: practising the judgement -- how to review a manuscript, write an authorship policy, raise a concern -- belongs to the site's peer-review and scientific-ethics module, not to this hour. Set expectations early: this is not a lecture about a scandal. There isn't one.

Be precise about what this slide claims. It does not say H01 lacks ethics approval -- the approval details are in the supplementary methods. It says the main article does not carry them, so a reader who stops at the article or the landing page has not seen them. The failure it warns against is citing without looking.

There is no single right answer to the follow-up, and that is intentional -- this is the seminar half of the lecture. The tension to surface: the same clinical detail that narrows the population is the detail a scientist needs to interpret the tissue (the pathology caveat two slides back).

Do not tidy the last two rows up. The hemibrain CC BY version and the H01 data licence could not be confirmed from a primary source, and the honest thing to teach is that a reuser must confirm both before redistributing either dataset. Presenting a guess as a licence is worse than presenting the gap.

Answer: the data licence, CC BY-NC 4.0. The paper being CC BY does not help; the product is built on FlyWire connectivity, not on the article. The decision table two slides on has the rows for this: "build a product or a paid service" (FlyWire's NC clause applies) and "train a model" (an NC clause reaches the model if the model is commercial).

Jwa & Poldrack is cited here for its spectrum finding only. Per-repository policy counts were not recovered from the primary source, so none are quoted.

The defensible phrasing uses only facts from the provenance slide in Part A. Point out that it is longer and less exciting, and that it is also the only version of the sentence that is true as written. If the room has seen the synapse-detection lecture, connect it: "the data says 74% excitatory" is a claim about a detector, and needs the same bounded-claim discipline.