Standards and reference desk

Working references, each annotated with when to use it. This is the canonical practical reference (the former docs/sources.md verify-status table is folded in here). When a tool’s behaviour is ambiguous, the standard/setool docs here are the source of truth.

Verification note: source families below were verified 2026-08-10. A specific claim still requires verifying the exact source at point of use, with a retrieval date. Never invent a citation.

---

DICOM

Resource	What	When to use it
DICOM Standard — dicom.nema.org/standard.html (VERIFIED 2026-08-10, 2026c ed.)	The normative standard (PS3.1–18). Key pages: Image Plane Module (geometry), Segmentation Image Module (SEG types).	Settle a precise question only. For geometry use PS3.3 Image Plane Module; for pixels use Image Pixel Module + Modality LUT (Rescale Slope/Intercept). Do not browse.
Innolitics DICOM browser — dicom.innolitics.com	Fast searchable browser over the standard	Default lookup: open this tab whenever a tag appears; read its definition, VR, module.
pydicom — pydicom.github.io	Python DICOM library + docs	Every hands-on task. “Dataset basics” and the “Pixel Data” tutorial cover ~90% of needs.
“DICOM is Easy” — dicomiseasy.blogspot.com	Gentle conceptual tutorial	First read-through of Patient→Study→Series→Instance and tag/VR/length from a real example.

Key parts to know exist (PS3.3): Image Plane Module (ImagePositionPatient, ImageOrientationPatient), Image Pixel Module (PixelData, BitsAllocated, RescaleSlope/Intercept), Modality LUT semantics, Multi-frame Functional Group macros (for SEG per-frame geometry).

Physical image handling

Resource	What	When to use it
SimpleITK fundamentals — simpleitk.org (read the fundamental concepts)	Origin/spacing/direction, physical vs index space, resampling	The single best reference for Ch. 2 geometry + resampling. Read the “fundamental concepts” page before any resampling code.
SimpleITK Resample / Transforms reference — docs	The Resample API, interpolators, transforms	When implementing image (BSpline/Linear) vs mask (NearestNeighbor) resampling.

Segmentation / visualization

Resource	What	When to use it
3D Slicer documentation — slicer.readthedocs.io	Loading studies, Segment Editor, volumes, transforms	Visual QC (Ch. 2 Part E), ROI inspection (Ch. 3). Read alongside the Witowski course + the 3D Slicer companion.
3D Slicer Training Compendium — training.slicer.org	Structured step-by-step tutorials with sample data	Companion exercises for Ch. 1–3 and learning/companion-3d-slicer.md.
DICOM SEG (Segmentation Storage; PS3.3 Segment Image IOD)	SEG object model: functional groups, source-image refs, segment identity, BINARY/FRACTIONAL/LABELMAP types	When decoding SEGs or reasoning about SEG↔mask round-trip (Ch. 2 Part C).
dcmqi / highdicom — github.com/QIICR/dcmqi / highdicom.readthedocs.io	SEG encoding/decoding beyond pydicom	When you must read/write SEGs robustly (segmentation metadata, fractional/labelmap types).
nnU-Net (Isensee et al., MICCAI 2021) — github.com/MIC-DKFZ/nnUNet	Self-configuring medical-image segmentation baseline	The strong default segmentation architecture (TotalSegmentator is built on it); read Ch. 6 Part C.
MONAI — monai.io	PyTorch medical-imaging DL framework + tutorials	2D/3D segmentation, transforms; integrates with Slicer (MONAIAuto3DSeg).

Quantitative imaging

Resource	What	When to use it
IBSI 1 — Zwanenburg et al., Radiology 2020;295(2):328-338 — PMID 32154773 / doi (VERIFIED); ref. manual arXiv:1612.07003; ibsi.readthedocs.io	Feature definitions, reference values, image-processing workflow	The standardisation layer for Ch. 3. Defines what a handcrafted feature is and how to process consistently.
IBSI 2 — Whybra et al., Radiology 2024 — doi (VERIFIED 2026-08-10)	Standardised convolutional filters (LoG/wavelet/Riesz) + reference filtered images/feature values	When you use filtered features — standardises filter semantics whose cross-software mismatch made filtered features hard to compare.
QIBA (Quantitative Imaging Biomarkers Alliance, RSNA) — rsna.org/QIBA; Profiles qibawiki.rsna.org (VERIFIED 2026-08-10)	Profiles with a performance Claim (bias + repeatability/precision under a stated protocol) for a specific biomarker	The broader measurement framework: FDG-PET SUV, DWI ADC, CT tumour volume change, etc. “What a measurement is worth.”
PyRadiomics — pyradiomics.readthedocs.io (features: features.html)	Open-source feature extraction (image type, binWidth, resampling, feature classes)	The implementation for Ch. 3. Pin parameters in a YAML file (as TRACE-CT does). Implements IBSI-style features with documented deviations — read the per-feature notes when strict cross-software reproducibility matters.
SlicerRadiomics — 3D Slicer extension	PyRadiomics in a UI	Build intuition interactively before scripting batch extraction.

Relationship: QIBA characterises the measurement (bias/precision) → IBSI 1/2 define the feature/filter computation → PyRadiomics implements → SlicerRadiomics visualises.

Agreement statistics, bootstrap, and robustness method

Resource	What	When to use it
Lin CCC — Lin LI, Biometrics 1989;45(1):255-268 — PMID 2720055 / doi (VERIFIED 2026-08-20)	The concordance correlation coefficient: agreement against the 45° line of identity (precision × accuracy)	Test–retest/reconstruction agreement of measurements. Not Pearson r — correlation ≠ agreement (Ch. 3 lesson 11).
Pearson / covariance / moments — Lane DM, Online Statistics Education (open textbook, VERIFIED 2026-08-20); ISLP Ch. 2	Correlation as standardised covariance; population vs sample moments	The first-principles layer under CCC; note 1/N vs N−1 changes CCC (not r).
Bootstrap — Efron B, Ann Statist 1979;7(1):1-26 — doi (Project Euclid, VERIFIED 2026-08-20); NIST e-Handbook §1.3 (bootstrap plots)	Resampling-with-replacement uncertainty estimation; percentile intervals	CI for statistics without analytic sampling distributions (e.g. CCC); sampling unit = subject pair in test–retest (Ch. 3 lesson 12).
ComBat original — Johnson WE et al., Biostatistics 2007;8(1):118-127 — PMID 16632515 (VERIFIED 2026-08-20)	Empirical-Bayes batch-effect adjustment (microarray origin)	What ComBat is, before any radiomics use.
ComBat in imaging (guide) — Orlhac F et al., J Nucl Med 2022;63(2):172-179 — PMID 34531263 / doi (VERIFIED 2026-08-20; full text in Zotero)	Radiomics/PET harmonisation guide; when ComBat assumptions are not met	Before applying ComBat to any multi-centre feature table; measure raw robustness first.
RIDER Lung CT v3 — TCIA collection page, DOI 10.7937/K9/TCIA.2015.U1X8A5NR, CC BY 4.0 (VERIFIED 2026-08-20)	32 NSCLC patients, same-day repeat scans, six reconstructions per scan (all six are current v3 content; v2 held only 1.25L); same-day repeat design: Zhao et al., Radiology 2009;252(1):263-272, PMID 19561260	The public test–retest + reconstruction dataset behind TRACE-CT M1-09.
Pfaehler et al. — Phys Imaging Radiat Oncol 2021;20:69-75 — PMID 34816024 / doi (VERIFIED; full text inspected 2026-08-23)	Systematic review of radiomics repeatability/reproducibility: acquisition, reconstruction, segmentation, and interpolation affect stability; sensitivity is feature-dependent; reporting checklist	Use when interpreting feature-family summaries or making workflow-level robustness claims; do not infer a universal stable family.
Haarburger et al. — Scientific Reports 2020;10:12688 — doi (VERIFIED; full text inspected 2026-08-23)	CT radiomics under manual and probabilistic segmentation variability; identifies robust and fragile features within categories and across datasets	Use when isolating segmentation as one possible source of feature instability; it does not identify the cause of a full test–retest result by itself.
Ye et al. — BJR Open 2022;4(1):20210083 — PMID 36452056 / doi (VERIFIED; full text inspected 2026-08-23)	Statistical framework for multi-step quantitative-imaging measurements: repeatability conditions, between-subject variation, CCC, confidence intervals, and difference-versus-mean plots	Use when interpreting point agreement metrics relative to cohort spread and when choosing plots/uncertainty summaries alongside CCC.

Missing data and synthetic tabular data (M1-07/M1-08 method families)

Resource	What	When to use it
mice — van Buuren S, Groothuis-Oudshoorn K, J Stat Software 2011;45(3):1-67 — doi (open access, FULL TEXT INSPECTED 2026-08-20: FCS, m parallel streams, Rubin pooling, MAR untestability)	Multivariate imputation by chained equations; canonical multiple imputation (M datasets, Rubin pooling)	The framework reference. Distinguish from single-table proxies (sklearn IterativeImputer) — TRACE-CT labels its own use honestly (Ch. 4 lesson 9).
CatBoost — Prokhorenkova L et al., NeurIPS 2018 — arXiv:1706.09516 (FULL TEXT INSPECTED 2026-08-20: permutation-prefix ordered boosting; ordered TS with smoothing prior; unbiasedness claim conditional)	Unbiased gradient boosting: ordered boosting, ordered target statistics for categoricals	Strong bounded baseline for mixed-type tabular prediction/imputation.
CTGAN — Xu L et al., Modeling Tabular Data using Conditional GAN, NeurIPS 2019 — arXiv:1907.00503 (FULL TEXT INSPECTED 2026-08-20: VGM mode-specific normalization, log-frequency training-by-sampling, conditional vector)	Conditional GAN for tables: mode-specific normalization, training-by-sampling, conditional generation	Synthetic tabular augmentation; conditional vs generate-then-filter distinction (Ch. 4 lesson 10).
TabDDPM — Kotelnikov A et al., ICML 2023, PMLR 202:17564-17579 — arXiv:2209.15421 / PMLR (FULL TEXT INSPECTED 2026-08-20: ε-MSE + averaged categorical KL/VLB objective)	Diffusion for mixed-type tabular data (Gaussian + multinomial noising)	Diffusion alternative to GANs for tables; unconditional generation + subgroup selection semantics.

Standing cautions (from TRACE-CT M1-07/08 evidence): synthetic ≠ anonymised ≠ useful ≠ clinically valid; utility must be measured on untouched real held-out data (fidelity alone is insufficient); risk indicators (Gower NN, exact/near-duplicates) are indicators, never a privacy guarantee; report every seed, retain negatives.

Research reporting / validation

Resource	What	When to use it
CLAIM 2024 — Tejani et al., Radiol AI 2024;6(4):e240300 — doi (VERIFIED 2026-08-10)	Reporting of imaging-AI model studies (classification/segmentation/reconstruction). Not for radiomics/biomarker studies.	Ch. 7 lens for an imaging-AI model paper.
TRIPOD+AI — Collins et al. 2024 — tripod-statement.org (VERIFIED 2026-08-10)	Reporting of prediction-model studies (regression or ML)	Ch. 7 lens for a prediction model.
PROBAST+AI — Moons et al., BMJ 2025;388:e082505 — doi (VERIFIED 2026-08-10); extends/replaces PROBAST-2019	Risk-of-bias appraisal of prediction-model studies	Ch. 4/7: the systematic way to find leakage paths.
CLEAR — Kocak et al., Insights Imaging 2023;14:75 — PMID 37142815 / doi (VERIFIED 2026-08-10); + CLEAR-E3 (2024) + METRICS	The 58-item radiomics reporting checklist (ESR/EuSoMII-endorsed)	Ch. 3/7: the primary radiomics appraisal tool — not CLAIM.

Framework selection by study type (Ch. 7)

• Prediction-model study (regression or ML) → TRIPOD+AI (reporting) + PROBAST+AI (risk of bias).
• Imaging-AI model study (classification / segmentation / reconstruction) → CLAIM 2024.
• Imaging-biomarker / radiomics study → CLEAR (NOT CLAIM 2024, which excludes radiomics); IBSI 1/2 for feature/filter computation.

Other useful anchors

Resource	When to use it
IHE — Radiology TF + AI profiles (AIW-I Rev 1.1 TI 2020, AIR Rev 1.3 TI 2025, AIRA AI Result Assessment, TI 2025) — profiles.ihe.net (VERIFIED 2026-08-10; all Trial Implementation)	Imaging-informatics workflow & interoperability failure modes (Ch. 5); AIRA for post-deployment assessment (Ch. 9).
RSNA / MICCAI / SIIM educational materials; TCIA datasets & challenges	Real data, methods, baselines (UNVERIFIED until used).
Bushberg, Essential Physics of Medical Imaging (recognised textbook)	CT acquisition/reconstruction/dose depth — read when dose/reconstruction becomes an active TRACE-CT variable.

Clinical bridge: interoperability, de-identification, governance, regulatory boundary (Ch. 5/9)

All rows below verified 2026-08-29 (full text / official text inspected unless noted); re-verify volatile items before relying on them in new claims.

Resource	What	When to use it	Status
DICOM PS3.2 — Conformance (§6, §7, Annex N; 2026c ed.)	Purpose/comparison claim of Conformance Statements; normative statement template (per-AE SOP Classes, roles, Transfer Syntaxes, options/limitations)	Interoperability-prediction questions (Ch. 5 lesson 9)	VERIFIED 2026-08-29
DICOM PS3.4 — Service Class Specs (2026c ed.): Annex B (Storage, levels of support/coercion), Annex J (Storage Commitment Push Model, N-ACTION/N-EVENT-REPORT), Annex F.7 (MPPS), Annex K (MWL)	Storage acceptance vs commitment responsibility; worklist/performed-step services	Storage Commitment and workflow-state questions (Ch. 5 lessons 1, 9)	VERIFIED 2026-08-29
DICOM PS3.15 Annex E — Attribute Confidentiality Profiles (2026c ed.)	Basic Application Level Confidentiality Profile: action codes D/Z/X/K/C/U, retention/cleaning options (incl. Clean Pixel Data, Clean Recognizable Visual Features, Retain Longitudinal Full Dates), Encrypted Attributes Data Set, Patient Identity Removed	De-identification profile design and appraisal (Ch. 5 lesson 10)	VERIFIED 2026-08-29
GDPR — Regulation (EU) 2016/679 (OJ L 119, 4.5.2016) — EUR-Lex	Art. 4(1) personal data; Art. 4(5) pseudonymisation; Recitals 26/28 (identifiability, anonymous data)	Legal anonymisation vs pseudonymisation boundary (Ch. 5 lesson 10)	VERIFIED 2026-08-29 (official text inspected via EUR-Lex)
EDPB — Opinion 28/2024 (17 Dec 2024, AI models/anonymisation, edpb.europa.eu); Guidelines 02/2026 on anonymisation (published 7 July 2026, public consultation draft)	Authoritative EU guidance on anonymisation as contextual, risk-based determination (singling out/linkability/inference)	De-identification claims touching GDPR	VERIFIED 2026-08-29 (identity + key content); Guidelines 02/2026 = draft, in consultation
MDR — Regulation (EU) 2017/745 (OJ L 117, 5.5.2017) — EUR-Lex	Art. 2 definitions (medical device incl. software; intended purpose), Recital 19 (software qualification)	Research-software vs clinical-product boundary (Ch. 9 lesson 10)	VERIFIED 2026-08-29 (official text inspected via EUR-Lex)
MDCG 2019-11 rev.1 — Qualification and classification of software (17 June 2025)	Endorsed (non-binding) guidance: intended-purpose-driven qualification; Rule 11 classification ladder	Device-software classification orientation (Ch. 9 lesson 10)	VERIFIED 2026-08-29 (official PDF inspected)
AI Act — Regulation (EU) 2024/1689 (OJ L, 12.7.2024) — EUR-Lex	Art. 6(1) high-risk = Annex I harmonisation legislation (MDR listed) + third-party conformity assessment; Art. 6(2)–(3) Annex III + derogations	AI-Act relationship to medical-device software (Ch. 9 lesson 10)	VERIFIED 2026-08-29 (official text inspected via EUR-Lex)
HL7 FHIR R5 — ServiceRequest, ImagingStudy, DiagnosticReport, Observation (current published version v5.0.0)	Verified resource definitions and imaging-workflow linkage (basedOn/study/result/derivedFrom; DICOM UID encoding)	Order→study→report→measurement bridge (Ch. 5 lesson 1, Ch. 5 reference)	VERIFIED 2026-08-29
FUTURE-AI — Lekadir K, et al. BMJ 2025;388:e081554, DOI 10.1136/bmj-2024-081554, PMID 39909534 (open access; erratum BMJ 2025;388:r340)	International consensus guideline: six principles (fairness, universality, traceability, usability, robustness, explainability), 30 lifecycle recommendations, stricter compliance expected for deployable tools	Trustworthy/deployable-AI design and governance questions (Ch. 6 lesson 8, Ch. 9 lesson 9)	VERIFIED 2026-08-29 (full text inspected)
Kitamura F, et al. “Teaching AI for Radiology Applications: A Multisociety-Recommended Syllabus from the AAPM, ACR, RSNA, and SIIM.” Radiol Artif Intell 2025;7(6):e250137, DOI 10.1148/ryai.250137 (open access; also co-published in Med Phys / JIIM)	Multisociety competency syllabus: User, Purchaser, Physician Collaborator, Developer personas (workflow, curation, bias/fairness, deployment, value, governance)	Coverage benchmark and role/competency consensus (Ch. 9 lessons 9–10) — never sole authority for normative claims	VERIFIED 2026-08-29 (full text inspected)

---

Rules

• Never invent a citation; never copy a citation without checking when the claim matters. Prefer DOI/PMID; record the retrieval date on verify.
• If reliable sources disagree, preserve the disagreement; do not silently resolve it by intuition.
• Distinguish and label: ESTABLISHED EVIDENCE · EXPERT CONSENSUS · COMMON PRACTICE · PLAUSIBLE INFERENCE · PROJECT HYPOTHESIS.