02 Brain Data Across Scales

Technical Training: Nanoscale Connectomics

Session outcomes (60 minutes)

• Match biological questions to minimal sufficient spatial/temporal scale.
• Choose representation transitions without losing inference-critical detail.
• Produce a scale-and-compute justification for one analysis plan.

Pedagogical arc

• Concept map: scale as an inference constraint.
• Modeling: question -> data product -> representation -> compute budget.
• Practice: learner selects scale stack for a case study.
• Check: defend tradeoffs under critique.

Why this matters

• Scale mismatch is a major source of invalid conclusions.
• Resolution, coverage, and compute are coupled design variables.
• "More data" does not fix wrong scale selection.

Visual context: multi-scale framing

• Instructor cue: ask what is visible here and what is fundamentally unobservable at this scale.

Visual context: analysis scale transition

• Distinguish acquisition scale from analysis target scale.

Visual context: representational conversion risk

• Volume -> segmentation -> skeleton/mesh -> graph can remove critical geometry.

Scale-selection framework

1. State estimand (what you will measure).
2. Determine smallest scale that resolves that estimand.
3. Verify coverage supports statistical claims.
4. Define acceptable uncertainty due to downsampling/registration.

Representation tradeoffs

• Raw volume: maximal fidelity, expensive queries.
• Segmentation: workable objects, boundary errors matter.
• Skeleton: topology-focused, diameter context reduced.
• Graph: fast analytics, spatial nuance largely removed.

Registration and uncertainty propagation

• Report transform model and residuals.
• Track uncertainty by region, not only global means.
• Carry registration confidence into downstream confidence intervals.

Compute realism for scale planning

• Storage and IO growth are nonlinear with resolution/coverage.
• Query latency determines practical iteration speed.
• Budgeting is part of scientific method feasibility.

Misconceptions to correct explicitly

• "Higher resolution always better."
• "Graph conversion is lossless enough for any question."
• "Registration error averages out automatically."

Think-Pair-Share (8 min)

Prompt: choose one hypothesis and argue for the minimum sufficient scale.

• Think: write one scale choice and one risk.
• Pair: challenge each other's coverage assumptions.
• Share: class votes on most defensible tradeoff.

In-class activity

Create a one-page scale plan:

• question,
• estimand,
• acquisition scale,
• analysis representation,
• compute/storage estimate,
• boundary statement.

Rubric checkpoint

• Pass: scale and representation choices are consistent with estimand.
• Strong: explicit uncertainty and compute tradeoff documented.
• Flag: claims exceed observable detail at chosen scale.

External paper figure slots (add in final teaching run)

• Kasthuri et al. 2015 (dense microcircuit reconstruction) figure on scale and completeness.
• MICrONS consortium overview figure on multimodal scaling.
• H01/human cortex connectomics figure on data scale and processing implications.

Bridge

Next unit: EM prep and imaging decisions that set artifact and quality limits.