Radiology Reporting

Radiology Reporting-Guideline Compliance

Use this skill to make an imaging study reviewer-proof on reporting. Radiology and the RSNA family require the relevant EQUATOR checklist at submission, and imaging-AI / radiomics papers are now judged against a specific, version-sensitive stack of guidelines. This skill (1) identifies the study type, (2) selects the correct guideline(s), (3) audits the manuscript item-by-item, and (4) returns a submission-ready checklist plus a prioritised fix list.

Core stance

• The checklist is the contract. A reviewer maps your paper to a guideline; do the same first, in their seat.
• Report honestly. Mark each item PRESENT, PARTIAL, or MISSING. Never label something compliant to be agreeable. A MISSING flag you surface is cheaper than a reviewer finding it.
• Cite the location. Every PRESENT claim must point to a section / page / figure / supplement. If you cannot point to it, it is PARTIAL at best.
• Versions matter. Use the current version (CLAIM 2024 Update, TRIPOD**+AI** 2024, CLEAR 2023, METRICS 2024). Name the version you audited against.
• Reporting ≠ quality ≠ risk-of-bias. CLEAR (reporting) → METRICS / RQS (methodological quality) → PROBAST(-AI) / QUADAS-2 (risk of bias). Different tools, different jobs; pick the right one(s).
• Don't invent the science. This skill audits reporting; it never fabricates the missing experiment, metric, or dataset. It tells the author what to add.
• Venue changes the stack, not the rigor. Radiology-family submissions stop at the guideline checklist; Nature-portfolio submissions add a Reporting Summary / Editorial Policy Checklist on top of the same guideline stack (→ nature-reporting-summary.md) — never treat the Reporting Summary as a replacement for CLAIM/TRIPOD+AI/CLEAR.

When to use

• "Which checklist does my study need?" / "What will Radiology require at submission?"
• "Audit this manuscript against CLAIM / TRIPOD+AI / STARD / CLEAR / METRICS / RQS."
• "Is my radiomics pipeline reported reproducibly (IBSI)?"
• "Fill in the CLAIM checklist with page numbers."
• "What's my risk-of-bias exposure under PROBAST-AI / QUADAS-2?"
• Pre-submission self-audit, or triaging a reviewer comment that cites a guideline.

Routing — pick the guideline(s) before auditing

Most imaging-AI papers need two or more of these (a reporting guideline and a quality/risk-of-bias tool).

Open references/guideline-router.md for the full decision tree, including hybrid studies (e.g. a radiomics prediction model validated for diagnostic accuracy → CLEAR + TRIPOD+AI + STARD + IBSI) and Nature-portfolio venues (add the Reporting Summary on top of whichever stack applies).

When to open extra files

Workflow

1. Classify the study. Determine task (classification / detection / segmentation / prediction / diagnostic accuracy / discovery), data provenance, whether a model is developed and/or validated, and whether the endpoint is accuracy, prognosis, or biology.
2. Select guideline(s) from the routing table. State which version. If the study is hybrid, select the stack and say why each applies.
3. Load the relevant reference file(s) and audit every item. For each item record: Item ID | Requirement (short) | Status (PRESENT/PARTIAL/MISSING/NA) | Location | Fix.
4. Prioritise fixes. Group into Blocker (will trigger major revision / desk reject), Should-fix (reviewer will likely ask), Polish. Tie each blocker to the specific reviewer risk.
5. Cross-check integrity hot-spots (see below) — the items reviewers weaponise most.
6. If the target is a Nature-portfolio venue, also complete the Reporting Summary / Editorial Policy Checklist (nature-reporting-summary.md) — additive, not a substitute.
7. Return the filled checklist + a one-screen executive summary + the prioritised fix list. Offer to draft the missing text/Methods sentences (hand off to radiology-writing).

Integrity hot-spots (audit these even if not asked)

These are the recurring reasons imaging-AI/radiomics papers get rejected:

• Data leakage / partition hygiene. Train/validation/test split made at the patient level (not slice/lesion); no test-set tuning; preprocessing, feature selection, harmonisation, and normalisation fit on training data only; augmentation never crosses the split. (CLAIM, TRIPOD+AI, METRICS, CLEAR all probe this.)
• External / independent validation. Internal CV alone is weak. State the validation type (internal resampling, temporal, geographic, fully external) and cohort source.
• Reference standard & ground truth. Who labelled, how many readers, expertise, blinding, adjudication, and the reference standard's own accuracy. (STARD, CLAIM.)
• Class/prevalence & spectrum. Report disease prevalence; flag artificial 1:1 sampling; describe the clinical spectrum (STARD spectrum bias; QUADAS-2 patient selection).
• Radiomics reproducibility. Software + version, image preprocessing (resampling, discretisation/bin width, intensity normalisation), segmentation method and inter-observer reproducibility (ICC), feature definitions IBSI-compliant, and scanner/protocol harmonisation (e.g. ComBat). (CLEAR, METRICS, IBSI.)
• Sample size / EPV. Events-per-variable, or a stated sample-size rationale (Riley et al. for prediction models). High-dimensional features vs. n is the classic overfitting trap.
• Metrics match the task & prevalence. AUC alone is insufficient; report calibration and clinical-utility (decision-curve) for prediction models; report CIs everywhere. (Hand off computation to radiology-stats.)
• Code / model / data availability. Statement present and specific. (Hand off to radiology-data.)

Output contract

Return, in this order:

1. Study classification — task, design, endpoint, and the selected guideline stack (with versions).
2. Checklist — a table with Item | Status | Location | Fix for every item of each selected guideline. Use NA only with a one-line justification.
3. Compliance summary — counts (PRESENT / PARTIAL / MISSING / NA) per guideline and an overall readiness read (e.g. "CLAIM 31/42 present; 4 blockers").
4. Prioritised fixes — Blocker / Should-fix / Polish, each tied to the reviewer risk and the manuscript location to edit.
5. Author input needed — questions only the authors can answer (e.g. "Was the test set sampled at patient level?").

If the user pastes only part of a manuscript, audit what is present and mark the rest Cannot assess — section not provided rather than guessing.

Quality bar

A good audit reads like a rigorous methods reviewer who is on the author's side: it finds the holes before submission, points to the exact item and location, and hands back the precise sentence the Methods needs — without ever inventing compliance the paper doesn't have.

Handoffs

• Missing statistics → radiology-stats (compute/report AUC CIs, DeLong, ICC, calibration, DCA).
• Missing Methods/Results prose → radiology-writing.
• Data/code availability wording, DICOM de-identification, Extended Data/Source Data → radiology-data.
• Radiogenomics-specific design/leakage → radiology-radiogenomics.
• Figure that proves an item (ROC, calibration, flow diagram) → radiology-figure.
• Explainability/uncertainty items for a DL model → radiology-deep-learning/interpretability-uncertainty.md.
• Checklist complete; want a full adversarial pre-submission read → radiology-prereview.