Design and audit a hand-crafted radiomics study end-to-end to Radiology (RSNA) / CLEAR / IBSI standard — image preprocessing (resampling, intensity normalisation, gray-level discretisation/bin width, filters), IBSI-compliant feature extraction (PyRadiomics or equivalent), reproducibility/stability filtering, leakage-safe feature selection, modelling, and internal/external validation. Use when the user plans or reviews a radiomics pipeline, mentions PyRadiomics, IBSI, feature extraction, bin width, gray-level discretisation, LASSO feature selection, radiomics signature/score, or "影像组学/放射组学". Produces a reproducible pipeline spec, runnable parameter settings, a leakage audit, and Methods text. Never fabricates feature counts or performance.
Design and audit a hand-crafted radiomics study end-to-end to Radiology (RSNA) / CLEAR / IBSI standard — image preprocessing (resampling, intensity normalisation, gray-level discretisation/bin width, filters), IBSI-compliant feature extraction (PyRadiomics or equivalent), reproducibility/stability filtering, leakage-safe feature selection, modelling, and internal/external validation. Use when the user plans or reviews a radiomics pipeline, mentions PyRadiomics, IBSI, feature extraction, bin width, gray-level discretisation, LASSO feature selection, radiomics signature/score, or "影像组学/放射组学". Produces a reproducible pipeline spec, runnable parameter settings, a leakage audit, and Methods text. Never fabricates feature counts or performance.
Hand-crafted Radiomics Study Design
Use this skill to build (or audit) a hand-crafted radiomics study that is reproducible and
leakage-free, from preprocessing through validation. Radiomics papers are desk-rejected for the
same recurring reasons: non-standardised features, segmentation not characterised, and data
leakage in selection/normalisation. This skill encodes the IBSI/CLEAR pipeline and the
partition hygiene reviewers enforce.
Core stance
IBSI or it isn't reproducible. Report image processing and feature definitions to IBSI
standard (resampling, discretisation, filters, aggregation, software+version) — otherwise
"feature X predicts Y" is irreproducible. (→ radiology-reporting/IBSI.)
Discretisation is a decision, not a default. Fixed bin width vs fixed bin count
changes every texture feature; state which, the value, and why; keep it consistent.
Segmentation error propagates. Use reproducible masks and filter unstable features
(ICC) before modelling (→ radiology-annotation).
Selection lives inside training only. Feature selection, normalisation, imputation, and
harmonisation are fit on training folds, never on the whole cohort — the classic leak.
Match complexity to events. Thousands of features vs tens of patients overfits; respect
EPV and validate honestly (→ radiology-stats).
Report calibration + utility, not just AUC, for a clinical signature (→ radiology-stats).
Integrity. Never invent feature counts, ICCs, or performance; mark what must be computed.
When to use
"Design / review my radiomics pipeline (PyRadiomics, IBSI)." / "影像组学流程设计或审查。"
"Bin width or bin count? what resampling/normalisation/filters?"
"How do I select features without leakage?" / "LASSO/mRMR feature selection 怎么做才不泄漏？"
"Build a radiomics signature/score and validate it."
"Audit this radiomics Methods for leakage and IBSI compliance."
Audit leakage (leakage-audit.md) and write Methods to CLEAR/IBSI.
Output contract
Pipeline spec — preprocessing → extraction → selection → model → validation, each step
with its parameters and the leakage control marked.
Parameters — resampling, normalisation, discretisation (bin width/count), filters,
feature families, software+version (a PyRadiomics parameter file where applicable).
Selection/modelling plan — method, where it sits relative to the split, EPV check.
Validation plan — internal + external; metrics incl. calibration/DCA.
Leakage audit — pass/fail per item with the fix.
Methods paragraph — CLEAR/IBSI-aligned prose (+ 待确认 for Chinese authors).
Quality bar
A good radiomics spec is one another lab could re-run from the parameters alone and get the same
features — with segmentation error quantified, selection inside the split, and performance
reported with calibration and CIs, never AUC alone.
