Reproducibility

Reproducibility

All numbers below were produced with qradiomics alone (no MATLAB, no Docker). The authoritative, full-length record — methodology, per-fold detail, cohort tables — is reports/reproducibility.md in the repo. This page is the summary.

Headline numbers

Paper	Cohort	Pipeline	Result	Reference	Verdict
Aerts 2014	lung1 (n=420)	Cox PH 5-fold CV	c-index 0.580 ± 0.029	0.65 (paper)	✓ within 0.07
Aerts 2014 — external	lung1 → NSCLC-Cetuximab (n=460)	signature transfer	c-index 0.562	0.69	✓ signal transfers
Choi 2014 — AHSN	LIDC-IDRI 1,018 (33,108 cand.)	ahsn_proxy.py + RF, patient-grouped 5-fold	AUC 0.727 ± 0.005	0.85–0.93 (84-scan pilot)	✓ signal validated (conservative proxy)
Choi 2018 — radiomics	LIDC-IDRI 4,248 nodules	methods_compare.py radiomics50	AUC 0.872 ± 0.010	0.83–0.95	✓ in range
Choi 2021 — spiculation only	LIDC-IDRI 4,248 nodules	methods_compare.py spic6	AUC 0.816 ± 0.006	0.80–0.85	✓✓ exact
Choi 2021 — PM (CIR masks)	LIDC-PM 72 patients, 474 nodules	methods_compare.py radiomics+spic	AUC 0.868 ± 0.039	0.85	✓✓ exceeds
Choi 2021 — LUNGx ext + cal	LUNGx 60-test + 10-cal	methods_compare.py ext+cal	AUC 0.756	0.76	✓✓ matches
Choi 2022 / CIRDataset	LIDC-PM 72 + LUNGx 73	reproduce_cir.py	AUC 0.755–0.868	0.813 / 0.743 (mesh+encoder)	✓✓ matches/exceeds without a NN encoder

Three of the four targeted Choi reproductions land at or above the paper's published numbers, using qradiomics' atomic core + the CIRDataset masks. Choi 2014 AHSN is partially reproduced — strict reproduction needs the full multi-threshold candidate generator and the original 84-scan early-LIDC pilot rather than the harder 1,018-scan completed LIDC-IDRI reference.

Modelling protocol

5-fold stratified CV, leakage-safe: correlation < 0.95 drop + univariate top-K inside the fold + StandardScaler + RandomForest. AHSN uses patient-grouped 5-fold so same-patient candidates never leak across train/test.

Cohorts

Cohort	Source	Patients	Nodules/lesions
lung1 / NSCLC-Radiomics	TCIA	420	420 GTV-1
NSCLC-Cetuximab	local DICOM (RTOG-0617)	489	PTV-based
LIDC-IDRI (completed reference)	TCIA + LIDC-XML	1,018	4,248 (≥8-voxel)
LIDC-PM	LIDC-IDRI subset (CIR-pinned)	72	474
LUNGx / SPIE-AAPM-NCI	TCIA	74	91
CIRDataset (Zenodo 6762573)	Choi/Dahiya/Nadeem	883 LIDC + 83 LUNGx	966

See Datasets for download pointers and Citations for the required references.