radiology-annotation

• agents/openai.yaml
• README.md
• references/label-ontology-and-reference-standard.md
• references/lesion-selection.md
• references/longitudinal-lesion-and-exam-linkage.md
• references/mask-geometry.md
• references/noisy-weak-and-report-derived-labels.md
• references/reader-protocol.md
• references/reproducibility-qc.md
• tests/test_mask_geometry_semantics.py

SKILL.md

---
name: radiology-annotation
description: "Design/audit imaging truth, readers, ROI geometry, reproducibility and label noise; not model training."
---

# ROI / VOI / Mask Annotation SOP

Use this skill to make the **annotation** behind a radiomics/segmentation/radiogenomics study
defensible. Reviewers reject papers when the segmentation is a black box: unknown readers,
no reproducibility, masks that don't align with the images, or feature instability never tested.
This skill specifies the lesion-selection strategy, the reader protocol, the reproducibility
plan, and the geometric checks — and writes the Methods text.

## Core stance

- **Segmentation is a measurement, so characterise its error.** Report inter-/intra-observer
  reproducibility and propagate it (drop unstable features) — don't treat masks as ground truth.
- **Pre-specify the lesion-selection rule.** 2D vs 3D, whole-tumour vs largest-slice vs
  peritumoral ring vs sub-regional habitat vs multi-lesion handling — decided up front, applied
  uniformly, and justified by the biology and the endpoint.
- **Readers are part of the method.** Number, seniority, blinding to outcome, independent vs
  consensus, and the adjudication rule for disagreement all belong in Methods.
- **Geometry must be exact.** Mask and image must share spacing, origin, direction, and slice
  order; a one-voxel or flipped-axis mismatch silently corrupts every feature.
- **Reproducibility before modelling.** Filter to reproducible features (e.g. ICC threshold)
  **before** selection/modelling, on training data only — leakage hides here too.
- **Integrity.** Never invent ICC/Dice values, reader counts, or QC results; mark what must be
  measured.

## When to use

- "Design an ROI/mask annotation SOP / 帮我写标注 SOP（勾画规范）。"
- "Two radiologists contoured the lesions — how do I report agreement and select stable features?"
- "Whole-tumour or largest-slice? 2D or 3D? peritumoral? habitat?"
- "My masks don't line up with the images / DICOM-SEG/RTSTRUCT conversion issues."
- "Write the segmentation + reproducibility paragraph for Methods."

## When to open extra files

| File | Open when |
|---|---|
| [references/lesion-selection.md](references/lesion-selection.md) | Choosing 2D/3D, whole-tumour/largest-slice/peritumoral/habitat/multi-lesion; what the endpoint and biology imply |
| [references/reader-protocol.md](references/reader-protocol.md) | Reader number/seniority, blinding, independent vs consensus, adjudication, training set |
| [references/reproducibility-qc.md](references/reproducibility-qc.md) | Repeat-annotation design, ICC/Dice/Hausdorff, feature-stability filtering, sensitivity analysis |
| [references/mask-geometry.md](references/mask-geometry.md) | DICOM/NIfTI/DICOM-SEG/RTSTRUCT spacing/origin/direction/slice-order checks, conversion pitfalls |
| [references/label-ontology-and-reference-standard.md](references/label-ontology-and-reference-standard.md) | Defining case/lesion/exam labels, indeterminate states, composite reference standards, adjudication and provenance beyond masks |
| [references/noisy-weak-and-report-derived-labels.md](references/noisy-weak-and-report-derived-labels.md) | Using report-mined, weak, noisy or incomplete labels without silently treating them as verified truth |
| [references/longitudinal-lesion-and-exam-linkage.md](references/longitudinal-lesion-and-exam-linkage.md) | Linking patients, examinations and lesions over time; new/disappearing/split/merged lesions and treatment-time ambiguity |

## Workflow

1. **Fix the target construct and reference standard.** Define case/lesion/exam ontology,
   indeterminate/failure categories, label source, verification window, adjudication and provenance;
   use the noisy/weak-label and longitudinal-linkage references when applicable.
2. **Fix the segmentation target.** What is segmented (whole tumour, core, necrosis, edema, node, organ),
   in which modality/sequence/phase, and why that target serves the endpoint.
3. **Choose the selection strategy** (lesion-selection.md) — dimension, region, peritumoral
   extent, habitats, and the multi-lesion rule. Justify it.
4. **Specify the reader protocol** (reader-protocol.md) — who, how many, blinded to what,
   independent vs consensus, adjudication, and any segmentation-training phase.
5. **Plan reproducibility** (reproducibility-qc.md) — repeat-annotation subset, ICC/Dice/HD,
   the feature-stability filter and threshold, and a sensitivity analysis.
6. **Verify geometry** (mask-geometry.md) — confirm image↔mask spacing/origin/direction/slice
   order; document the conversion path (DICOM-SEG/RTSTRUCT/NIfTI) and QC.
7. **Write Methods** — the annotation paragraph: target, software+version, readers, blinding,
   reproducibility, stability filtering — in _Radiology_ style.

## Output contract

1. **`Annotation SOP`** — target, selection strategy, software/version, step-by-step procedure.
2. **`Ground-truth passport`** — ontology, unit, source, verification window, indeterminate/failure
   handling, adjudication, label confidence and patient–exam–lesion linkage.
3. **`Reader protocol`** — number/seniority/blinding/independent-or-consensus/adjudication.
4. **`Reproducibility & QC plan`** — repeat design, metrics (ICC/Dice/HD), stability filter +
   threshold (frozen here from measured reproducibility, applied by `radiology-radiomics` inside
   training), sensitivity analysis — with placeholders for values to be measured.
5. **`Geometry checklist`** — the alignment/conversion checks and their pass/fail.
6. **`Annotation provenance`** — mask/image IDs or irreversible hashes, version, reader,
   timestamp/status, adjudication/revision history, software/version, and geometry-QC artifact
   when a full project record is requested.
7. **`Methods paragraph`** — submission-ready prose (+ 待确认 list for Chinese authors).

## Quality bar

A good annotation spec lets an independent group reproduce the masks and the feature set, tells
the reviewer exactly how segmentation error was measured and controlled, and never reports an
agreement value that was not computed.

## Handoffs

- Series/phase/sequence acceptance and image-formation/QC passport → `radiology-acquisition-qc`;
  annotation owns labels, contours, reader process and image-mask geometry, not scanner conformance.
- IBSI feature reproducibility / CLEAR-METRICS audit → `radiology-reporting`.
- Hand-crafted feature extraction pipeline → `radiology-radiomics`.
- ICC model choice, Dice/Bland-Altman statistics → `radiology-stats`.
- Habitat masks for imaging×omics → `radiology-radiogenomics`.
- De-identifying shared masks/images → `radiology-data` / `radiology-ethics`.
- Project manifest and downstream staleness after mask revisions → `radiology-pipeline`.
- This skill specifies and reports annotation; it does not perform clinical contouring.