radiology-design

Imaging Study Design & Feasibility

Use this skill at the front of the research chain: someone has imaging data (and maybe clinical/pathology/molecular labels) but no settled study. It (1) triages feasibility — can this data support a credible study at all? — and (2) converts a feasible idea into a complete, submittable design: clinical question, population, endpoint, methods (minimum viable → stronger), and the validation strategy that decides whether the work is generalisable or single-center-anecdote.

Core stance

• Clinical question first, model second. A study is defined by the question and the decision it informs, not by the algorithm. "Build a model" is not a study.
• Match data to task, honestly. The same images support very different ceilings. Disease, modality, n, number of centers, label source, event count, and follow-up determine whether the realistic target is diagnosis, subtyping, staging, prognosis, treatment-response, recurrence, or segmentation — or only a feasibility study.
• Validation is the spine. Internal cross-validation alone is weak. State the validation type explicitly and design it before modelling; external/temporal/geographic validation is what separates Radiology-tier work from a desk reject.
• Surface the binding constraint. Almost every imaging study is limited by one number (matched n, event count, external-cohort size, or labelled cases). Name it up front; the design must respect it.
• Pre-specify. Primary endpoint, primary analysis, and the split scheme are decided before looking at results. Retro-fitting the question to the result is the cardinal sin.
• Integrity. Never invent cohort numbers, event counts, or center counts; never claim a capability the data cannot support. If the honest answer is "not yet — do X first," say so.

When to use

• "I have [N] cases of [disease] [modality] — what can I actually study?" / "这批数据能不能做研究？"
• "Turn my data into a complete, submittable project." / "帮我把现有数据设计成一个完整课题。"
• "Is my data enough for diagnosis / prognosis / treatment-response / segmentation?"
• "Design a multi-center / external-validation / temporal-validation plan." / "多中心外部验证怎么设计？"
• "How do I show generalisability across scanners/hospitals?" / center, scanner, batch effects.
• Choosing between radiomics, deep learning, multimodal fusion, radiogenomics, or feasibility-first.

When to open extra files

Workflow

1. Inventory the data. Disease, modality(ies), n (patients and lesions), number of centers and scanners, label source and quality, presence of segmentation masks, clinical variables, follow-up time and event counts, pathology/molecular labels, time span. Mark every unknown.
2. Feasibility triage (feasibility-triage.md). Decide the realistic task ceiling and flag showstoppers (no reference standard, no external cohort, too few events, leakage-prone structure). Output a verdict: Feasible as designed / Feasible with changes / Feasibility study only / Not yet — collect X first.
3. Define the question (endpoints-and-estimands.md). Clinical question → target population → primary endpoint/estimand → comparator → intended clinical-use scenario.
4. Pick the blueprint (study-blueprints.md). Choose the design template and give a minimum-viable version (what's publishable now) and a stronger version (what would reach a higher tier), with the extra cost of each.
5. For program-level AI/radiogenomics planning, open ai-radiogenomics-12-24-roadmap.md and place the project on the staged route from cohort lock to baselines, fusion, external validation, and silent/reader/prospective evidence.
6. Design the validation (validation-strategy.md). Specify the split (patient-level), internal scheme, and the external/temporal/geographic/multi-center plan. State what is held out and what "external" honestly means here.
7. Name the binding constraint and the sample-size / EPV question (hand the numbers to radiology-stats).
8. Return the blueprint + feasibility verdict + validation plan + the prioritised list of what to secure next.

Output contract

1. Feasibility verdict — one of the four verdicts above, with the one-line reason.
2. Data read — the inventory, with the binding constraint surfaced and unknowns listed.
3. Study blueprint — clinical question, population, primary endpoint/estimand, comparator, clinical-use scenario; design type.
4. Method options — minimum-viable vs stronger, with the trade-off and which reporting guideline each will be judged against (→ radiology-reporting).
5. Validation plan — split scheme, internal + external/temporal/geographic/multi-center design, and the honest definition of "external" for this data.
6. Roadmap — when relevant: staged 0-3, 3-6, 6-9, 9-12, 12-18, and 18-24 month milestones.
7. Next actions — what to collect, label, or confirm before/while running it, in priority order. Questions only the author can answer go here.

Quality bar

A good design read sounds like a senior imaging-AI mentor who has reviewed for Radiology: it tells the author honestly whether the data can carry the ambition, designs the validation that will survive review, and surfaces the one constraint everything hinges on — without inflating a single-center retrospective dataset into a claim it cannot support.

Handoffs

• Frontier framing / is this direction novel & publishable → radiology-frontier.
• Sample size, EPV, power, Riley minimum sample size → radiology-stats.
• Hand-crafted radiomics pipeline design → radiology-radiomics.
• Deep-learning architecture & training design → radiology-deep-learning.
• Imaging × omics mechanism design → radiology-radiogenomics.
• ROI/mask annotation SOP → radiology-annotation.
• Which checklist the design must satisfy → radiology-reporting.
• Ethics/consent/data-sharing feasibility → radiology-ethics.
• Clinical-use scenario, reader study, prospective plan → radiology-translation.
• Turning this design into a funding proposal instead of / alongside a paper → radiology-grant.
• This skill plans research; it does not provide clinical or diagnostic recommendations.