08 Segmentation and Proofreading

Technical Training: Nanoscale Connectomics

Unit 08 · Segmentation and proofreading

Session outcomes (60 minutes)

  • Classify merge, split, boundary, and identity errors reproducibly.
  • Prioritize corrections by expected scientific impact.
  • Connect proofreading actions to quantitative QC metrics.
Unit 08 · Segmentation and proofreading

Pedagogical arc

  • Model: error taxonomy + live correction logic.
  • Practice: triage and correction on mixed cases.
  • Consensus: adjudicate borderline errors.
  • Check: correction log with metric rationale.
Unit 08 · Segmentation and proofreading

Why proofreading is scientific QC

  • Correction policy determines analysis validity.
Unit 08 · Segmentation and proofreading

Error taxonomy visual

  • Enforce explicit error-class coding in logs.
Unit 08 · Segmentation and proofreading

Ultrastructure-informed correction

Unit 08 · Segmentation and proofreading

The merge/split dial, on one real object

  • H01 ships two agglomerations of the same data.
  • c2 calls this one object; c3 calls it two. Neither is a bug.
  • Across 104 proofread cells: c3 needed 1.6× fewer merges, 2.1× more splits.
Unit 08 · Segmentation and proofreading

Before and after a human fixed it

  • Green = genuinely this cell. Red = 11,038 voxels the algorithm wrongly absorbed.
  • The membrane it crossed is real but faint — would you have caught it?
  • Note where the error is NOT: at the soma, automated needs no correction.
Unit 08 · Segmentation and proofreading

Synapse-aware correction checks

Unit 08 · Segmentation and proofreading

Organelle-assisted disambiguation

Unit 08 · Segmentation and proofreading

Boundary failure case

  • Show when to stop and escalate instead of over-correcting.
Unit 08 · Segmentation and proofreading

Identity-sensitive correction context

Unit 08 · Segmentation and proofreading

Metrics and release gates

  • VI
  • edge precision/recall
  • ERL
  • synapse-centric F1
Unit 08 · Segmentation and proofreading

Operational proofreading loop

  1. Triage by expected downstream impact.
  2. Correct with local + global consistency checks.
  3. Update targeted metrics.
  4. Route unresolved cases for adjudication.
  5. Gate release on predefined thresholds.
Unit 08 · Segmentation and proofreading

Misconceptions to correct

  • "Fix easiest errors first."
  • "Global metric improvements guarantee biological validity."
  • "Automation removes need for human policy."
Unit 08 · Segmentation and proofreading

Activity

Submit one correction log containing:

  • error class,
  • before/after rationale,
  • metric impact expectation,
  • confidence and escalation status.
Unit 08 · Segmentation and proofreading

Rubric checkpoint

  • Pass: correction decision tied to error class and metric logic.
  • Strong: priority ranking aligned to scientific impact.
  • Flag: edits without audit trail or rationale.
Unit 08 · Segmentation and proofreading

External paper figure integration

  • Januszewski et al. 2018 FFN architecture/performance figures.
  • Segmentation benchmark (CREMI-style) error/metric figures.
  • Human-machine proofreading workflow figures from open platforms.
Unit 08 · Segmentation and proofreading

External inserted figure pair (open license)


Unit 08 · Segmentation and proofreading

References and attribution

Unit 08 · Segmentation and proofreading