Skip to content
Imaging Evidence Path

Literature map

The intellectual / navigation layer over the literature. Zotero remains the canonical PDF/bibliography store; this map lets you jump to any topic and read intelligently now — one click from the source. It imposes no curriculum and no global reading cap.

Roles per item: Core · Worth reading · Go deeper · Watch · Historical. PMIDs/DOIs verified 2026-08-10 (original set) and 2026-08-22 (radiomics-expansion set) unless marked UNVERIFIED; peer-reviewed vs preprint is marked on each. No scientific claim in the chapters depends on an UNVERIFIED item.

CT & DICOM (Ch. 1–2) — tools/standards (no papers)

Geometry / SEG / resampling (Ch. 2)

  • SimpleITK fundamentals — simpleitk.org (read the “fundamental concepts” page) — Core tool. · dcmqi / highdicom — SEG encode/decode — Go deeper.

3D Slicer (companion + Ch. 2–3)

  • Jan Witowski — 3D Slicer Tutorial (11-video playlist) — YouTube playlist — Core course (see learning/companion-3d-slicer.md).
  • 3D Slicer Training Compendium — training.slicer.org — Core (official tutorials + data).
  • SlicerRadiomics extension — Worth reading (interactive radiomics).

Quantitative imaging & radiomics (Ch. 3)

  • IBSI 1 — Zwanenburg et al., Radiology 2020;295(2):328-338 — PMID 32154773 / doi — Core (feature definitions + reference values).
  • IBSI 2 — Whybra et al., Radiology 2024 — doi — Core (standardised LoG/wavelet/Riesz filters).
  • PyRadiomics — pyradiomics.readthedocs.io (features: pyradiomics.readthedocs.io/en/latest/features.html) — Core tool (IBSI-style, with documented deviations).
  • QIBA (RSNA) — rsna.org/QIBA · Profiles qibawiki.rsna.org — Core framework (measurement bias/repeatability Claims).
  • CLEAR — Kocak et al., Insights Imaging 2023;14:75 — PMID 37142815 / doi — Core (appraisal) (58-item radiomics checklist; +CLEAR-E3, METRICS).
  • PySERA — Salmanpour et al., Comput Methods Programs Biomed 2026:109463 — PMID 42235274 — Worth reading (tool) (557 features incl. 487 IBSI-compliant; >94% IBSI-reference reproducibility, near MITK; DL embeddings included; author-caveated tool comparison — peer-reviewed abstract + author-accepted manuscript inspected 2026-08-22; dataset-accuracy figures differ between preprint and journal abstracts and are not cited).
  • Robustness filtering ↔ predictive utility (mixed evidence; Ch. 3 lesson 06) — GBM segmentation-robustness trade-off: arXiv:2607.23626 (preprint, full text inspected — 4,752 features, 48.1% robust at ICC ≥ 0.85, filtering no detectable gain, model-selected features less robust; single retrospective cohort, no external validation) · inter-reader reliability + survival: Müller-Franzes G et al., Diagnostics 2022;12(2):247 — PMID 35204338 / doi (peer-reviewed, full text inspected — small gains: C-index 0.58 vs 0.56 CT, non-significant MRI) · stability filtering improved IDH prediction across segmentation schemes: AJNR 2026 — ajnr.A9229 (peer-reviewed, abstract inspected).

Negative controls / volume confounding (Ch. 3 falsification)

  • Welch et al. — “Vulnerabilities of radiomic signature development: the need for safeguards,” Radiother Oncol 2019;130:2-9 — PMID 30416044 / doi — Core (peer-reviewed; voxel-randomised reproduction of an Aerts-style signature: c-index 0.64 real = 0.64/0.60 randomised = 0.64 volume-only; three of four features volume-correlated).
  • READII-2-ROQC — “Negative controls reveal volume-driven confounding in radiomics and imaging foundation model features,” arXiv:2607.28423 (2026) — Watch (preprint; 9 volume-preserving control configurations over 3,552 tumours; Aerts signature stable under all controls; FMCIB features partly background-driven; open-source framework github.com/bhklab/readii_2_roqc; full text inspected 2026-08-22).
  • Aerts et al. — the original four-feature signature, Nat Commun 2014;5:4006 — doi — Historical/Core context (read with Welch: the original reports signature > volume; later controls show largely volume-driven behaviour under original definitions — disagreement preserved, not silently resolved).
  • Hatt et al. — “Radiomics: Data Are Also Images,” J Nucl Med 2019;60(Suppl 2):40S-44S — doi — Worth reading (volume/SUV confounding; definition changes can restore complementary value; PET-centred).
  • Ger et al. — PLOS ONE 2019;14(9):e0222509 — doi — Go deeper (volume+HPV beat radiomics-augmented models in H&N; controlled-protocol subset). · Fave et al. — Transl Cancer Res 2016;5(4):349-363 — doi — Go deeper (preprocessing changes feature/volume dependence).

Region definition (Ch. 3 intra-/peri-/multi-region)

  • Chen, Xu & He — Ki-67 in T1 NSCLC, tumour + 3 mm peritumoral best, J Thorac Dis 2026;18(2):67 — doi — Worth reading (multicentre, TRIPOD-reported; full text inspected).
  • Frontiers Oncol 2025 — intratumoral/peritumoral (3/6/9 mm) + 3D DL for NSCLC OS — doi — Worth reading (6 mm selected — contrast with Chen’s 3 mm: ring-width “optima” are study-specific).
  • BMC Cancer 2025 — induction-therapy response, 3 mm dilation minus tumour with chest wall/vertebral exclusion — doi — Go deeper (anatomical ring clipping as part of the measurement).
  • Multi-region / dose-region radiomics — no fully verified anchor yet; the best-matching candidate (multi-region radiation-pneumonitis model, external validation claimed) is not yet full-text verified — see verification queue item 2. The concept is taught without study-level claims in Ch. 3 lesson 15.

Delta / longitudinal radiomics (Ch. 3)

  • De Robertis et al. — prospective NALIRIFOX PDAC delta study, Transl Oncol 2026 — doi — Worth reading (107→55; ICC≥0.90 filtering; bootstrap; delta AUC 0.85 vs 0.76/0.82; Δ-volume strongest predictor — the ladder lesson applies to deltas; open-access full text inspected).
  • Wang et al. — externally validated PDAC delta radiomics, Ann Surg Oncol 2024;31:2608-2620 — PMID 38151623 / doi — Worth reading (discovery + external validation; combined clinical+delta best; C-index ≈0.66 external OS).
  • Duan et al. — CT quantitative imaging in PDAC after neoadjuvant therapy (narrative review), Gland Surg 2026;15:201 — PMID 42591759 / doi — Worth reading (delta features as preliminary-tier evidence; comparability/registration/stability requirements).

Opportunistic / pre-diagnostic imaging (Ch. 3 question families)

  • Korfiatis et al. — AI detection of visually occult PDA on prediagnostic CT, Gastroenterology 2023 — PMID 37657758 — Worth reading (abstract inspected; sensitivity 0.75 at median 475-day lead time; prospective validation still warranted).
  • ePAI — early/prediagnostic PDAC detection from CT, arXiv:2601.22134 — Watch (preprint; 75/159 prediagnostic patients detected, median 347-day lead; 30-radiologist multi-reader comparison; external 6-centre test).
  • Body-composition phenotype example (already verified): Ziegelmayer et al., Lancet Reg Health Eur 2025 — see EuSoMII section below.

Statistical learning & validation (Ch. 4)

  • Tran et al. — DL ACL-tear knee MRI, multi-continental external validation, Eur Radiol 2022;32(12):8394-8403 — PMID 35726103 / doi — Core (Ch. 7 worked example). Full-text lesson: the abstract’s headline external AUCs (0.962/0.922) are retrained/fine-tuned; the frozen-travel AUCs are 0.941/0.860, and thresholds were re-estimated per dataset — so “external validation” ≠ “frozen model travels.”
  • PROBAST+AI — Moons et al., BMJ 2025;388:e082505 — doi — Core framework (prediction-model risk of bias).
  • TRIPOD+AI — Collins et al. 2024 — tripod-statement.org — Core framework (prediction-model reporting).
  • ISLP / StanfordOnline Statistical Learning with Python — Core course (Ch. 4 is the imaging companion).

Imaging informatics (Ch. 5)

Medical-imaging AI (Ch. 6)

  • TotalSegmentator CT — Wasserthal et al., Radiol AI 2023;5(5):e230024 — PMID 37795137 / doi — Core (operational segmentation tool; built on nnU-Net).
  • TotalSegmentator MRI — Akinci D’Antonoli et al., Radiology 2025;314(2):e241613 — PMID 39964271 / doi — Worth reading (MRI seg behaviour).
  • MRSegmentator (whole-body MRI/CT) — Häntze et al., Radiol Artif Intell 2025;7(6):e240777 — PMID 40767616 / doi — Worth reading (cross-modality 40-structure seg; external testing on NAKO/AMOS/TotalSeg-MRI).
  • nnU-Net — Isensee et al. (MICCAI 2021) — Go deeper (self-configuring baseline).
  • MONAI — monai.io — Worth reading (tool/framework).
  • CLAIM 2024 — Tejani et al., Radiol AI 2024;6(4):e240300 — doi — Core framework (imaging-AI model studies; not radiomics).
  • Merlin — CT VLM foundation model, Nature 2026 — PMID 41781626 / doi (preprint PMID 38978576) — Worth reading (3D VLM; foundation-model evaluation at scale).
  • a2z-1 — Rajpurkar et al. — arXiv:2412.12629 — Worth reading (preprint; generalist abd/pelvis CT, strong external validation).
  • OmniMRI — He et al. — arXiv:2508.17524 — Watch (preprint; unified MRI VLM: recon/seg/detect/diagnose/report).
  • Foundation-model reviews — Akinci D’Antonoli et al., Diagn Interv Radiol 2026;32(3):259-272 — PMID 40626693 / doi — Worth reading (fundamentals/applications/risks; distinct from Paschali). · Paschali et al., Radiology 2025;314(2):e240597 — PMID 39903075 / doi — Worth reading (“what/how/why/why-not” orientation).
  • GRAI (generalist radiology AI) — Dogra et al., Radiology 2025;316(3) — doi — Watch (review/perspective proposing GRAI; financial/operational/clinical framing).
  • GPT-4 protocol selection — Gertz et al., Radiology 2023;307(5):e230877 — PMID 37310247 / doi — Watch (feasibility of LLM study/protocol determination from request forms; on a protocol-routing task).

Representation comparison: handcrafted vs learned (Ch. 6)

  • Foundation models vs radiomics benchmark — Neukirch, Maurer & Strodthoff, arXiv:2607.01001 (2026) — Worth reading (preprint; extractors × classifier heads × segmentation regimes, LUNG1→LUNG2, worst-case cross-cohort metric; task-dependent dominance; radiomics competitive for survival; label-derivation artifact named; full text inspected 2026-08-22).
  • VPI multi-component model — Insights Imaging 2026;17:194 — PMID 42498861 / doi — Worth reading (radiomics vs ViT vs combined; combined best in training only — externally 0.844/0.823/0.816; combined wins sensitivity, not AUC; full text + supplement inspected).
  • Upadhaya et al. — foundational AI vs radiomics vs clinical on RTOG 0617, Br J Radiol Open 2023 — doi — Worth reading (≈equal AUCs; three-way ensemble no gain).
  • Li et al. — handcrafted + DL features across three lung ROIs (pneumonitis), Radiat Oncol 2026 — doi — Go deeper (region × representation interaction; hybrid best internally, 0.828 external).
  • “Big, Bright, or Invisible” — frozen 3D CT encoder benchmark, arXiv:2608.05960 — Watch (preprint; no universal SOTA; detectability scales with contrast × spatial extent — a physical bottleneck for pooled embeddings).

Synthetic imaging (Ch. 6 Part E; Ch. 8 Part D) — verified representative set

  • MR→synthetic CT (review) — Spadea et al., “Deep learning based synthetic-CT generation in radiotherapy and PET: A review,” Med Phys 2021;48(11):6537-6566 — PMID 34407209 / doi — Worth reading (problem family B: MR→CT for electron density/dose; clinical context).
  • Reconstruction / denoising — Sim et al. (shoulder MRI DL-recon), Cureus 2025;17(10):e94561 — PMID 41246632 / doi — Worth reading (problem C; shoulder, not elbow). · Park et al. (lumbar CT DL denoising), Eur Radiol 2025;35(12):7867-7876 — PMID 40483292 / doi — Worth reading (problem C). · Faster Elbow MRI DL recon — Herrmann et al., Diagnostics 2023;13(17):2747 — PMID 37685285 / doi — Worth reading (distinct from PMID 41246632).
  • Choi et al. — synthetic contrast-enhanced CT from non-contrast chest CT (GAN, observer study, internal-only), Sci Rep 2021;11:20403 — doi — Worth reading (the strong-evaluation pattern shape; full text inspected 2026-08-22).
  • Task-based evaluation (perceptual realism insufficient) — “RaD: A Metric for Medical Image Distribution Comparison,” arXiv:2412.01496 — Worth reading (preprint; problem D — shows upstream perceptual metrics like FID do not track downstream segmentation performance; T1/T2 MRI translation).
  • Lecture example unresolved — the specific webinar T2→T1/STIR sequence-synthesis paper could not be reliably identified; do not conflate it with another paper. The method family (sequence-to-sequence translation) is represented by the MR→CT review and the RaD paper above.

Clinical deployment / lifecycle (Ch. 9)

  • Herpe et al. — AI assistance in radiologist knee-MRI interpretation, Eur Radiol 2026;36(2):1294-1305 — PMID 40745051 / doi — Worth reading (Ch. 9 reader-study example; abstract level). Lesson: 6 readers × 165 cases, AI trained on a separate 23k-study set, consensus reference; AI improved accuracy and inter-reader agreement (κ 0.54→0.78) — a retrospective reader+AI demonstration, not deployment evidence.
  • Yang et al. — AI boost to MRI lumbar spine reporting, Eur J Radiol 2024;179:111636 — PMID 39133990 / doi — Watch (reporting/workflow informatics).
  • Song et al. — “Determinants of Reader-LLM Interaction in Thoracic Radiology: Impact of Model Confidence and Reader Expertise,” Radiology 2026;320(2):e253397 — doi — Worth reading (10 readers × 100 cases, GPT-5/GPT-4o; rationale quality double-edged, expertise protective — journal abstract inspected only; full text paywalled and pending; effect sizes deliberately not cited in lessons).
  • LLM explanation format RCT — “The effect of medical explanations from large language models on diagnostic accuracy in radiology,” npj Digit Med 2026 — PMID 42026140 — Worth reading (2,020 assessments; chain-of-thought +12.2% accuracy; differential-diagnosis format up to 80% adherence when the LLM was wrong; full text inspected 2026-08-22).
  • “Care to Explain?” — local vs global AI explanations, chest radiographs, Radiology 2024 — PMID 39560483 — Go deeper (local explanations increased “simple trust” regardless of advice correctness; multicentre randomised).
  • ReCo (agentic/MCP framework; Orthanc+Merlin+TotalSegmentator+DosimeTron) — Tzanis & Klontzas — medRxiv 2026.07.14.26358025 / github.com/eltzanis/ReCo — Watch (preprint; compounding-error/provenance risks).

Research methodology (Ch. 7)

  • CLAIM 2024 (above) · TRIPOD+AI (above) · PROBAST+AI (above) · CLEAR (above) · IBSI 1/2 (above) — choose by study type (Ch. 7 table).
  • Tran et al. (above) — worked dissection vehicle.

EuSoMII 2026 MSK / webinar — historical/landscape

  • Ziegelmayer et al. — IMAT/lean muscle MRI in chronic back pain, Lancet Reg Health Eur 2025;54:101323 — PMID 40487774 / doi — Worth reading (segmentation→measurement→phenotype→clinical-question chain).
  • Ruitenbeek et al. — AI in MSK imaging: realistic applications, Skeletal Radiol 2024;53(9):1849-1868 — PMID 38902420 / doi — Worth reading (~1 h landscape map).
  • Guermazi et al. — Radiology 2024;310(1):e230764 — PMID 38165245 (erratum PMID 38289220) — Historical (MSK-AI trajectory; partly superseded).
  • Fritz & Fritz — Skeletal Radiol 2022;51(2):315-329 — PMID 34467424 · Gorelik et al. — Semin Musculoskelet Radiol 2020;24(1):38-49 — PMID 31991451 — Historical (older scoping/narrative reviews).

Unresolved

  • “Segmenting Whole-Body MRI and CT” webinar item — resolved as MRSegmentator (Häntze et al., PMID 40767616, above).

Verification queue — radiomics expansion (2026-08-22)

Daily-Briefing leads that could not yet be verified to the standard of .kilo/rules/10-evidence-and-sources.md. No learner-facing chapter depends on these. A failed verification is preserved honestly; re-check before promoting any item into a thematic section above.

  1. Registration-guided GAN for virtual contrast-enhanced CT — briefing 2026-08-22 described a Radiology: Cardiothoracic Imaging paper (published ≈2026-08-20): registration-guided GAN, non-contrast → virtual contrast-enhanced CT, internal + external testing, reader-performance component. State: not found — targeted searches (venue + “registration”
    • GAN + virtual contrast + reader, Aug 2026) surfaced adjacent work (Choi Sci Rep 2021; GCTANet, IEEE Access 2025; SC-BBDM diffusion bridge, arXiv:2508.16897; CycleGAN cardiac, BMC MIDM 2026) but none matching venue + design. Still to check: journal TOC / advance-publication listings for Radiol Cardiothorac Imaging around 2026-08-20; exact registration-guidance mechanism; whether the reader study was diagnostic-performance or realism-only. Supports: Ch. 6 lesson 6 (worked-pattern box currently anchored on Choi 2021, which is verified).
  2. Multi-institutional NSCLC tumour + peritumoral + lung + dose regions, external validation, high combined C-index — briefing lead. Best matching paper found: Kawahara D, Imano N, Kishi M, Fujiwara T, Kimura T, Murakami Y. “Prediction of radiation pneumonitis after CRT in patients with advanced NSCLC using multi-region radiomics and attention-based ensemble learning,” Med Phys 2025;52:e70140 — PMID 41261069 / doi:10.1002/mp.70140. State: identity probable, NOT full-text verified — journal abstract inspected (137 patients, two Japanese centres, 40 anatomical/dose- stratified regions incl. GTV/peritumoral shells/lung/dose subvolumes, four Radscores, external validation claimed, C-index 0.83 training / 0.83 external vs 0.54 clinical per abstract); full text is paywalled (no PMC copy), so cohort structure, region construction, train/external separation, and exact metrics remain unchecked — and the briefing may still describe a different 2026 paper. Removed from learner-facing support on 2026-08-22 for this reason; do not restore without full-text verification. Supports: Ch. 3 lesson 15 (multi-region bullet, currently taught without study-level claims).
  3. Cardiotoxicity radiomics and opportunistic CT bone-density examples (briefing leads) — not individually pursued; the Ch. 3 “question families” synthesis deliberately names these as families without study-level claims. Before adding any example: verify cohort, endpoint, and whether the marker was validated against a reference standard (DXA / strain / events), not just internally cross-validated.

How to use this map

  • Jump anywhere. Interested in Merlin? Open its row, then read a Paschali/Akinci D’Antonoli review alongside. No queue, no gate.
  • “Saved” ≠ “must read.” Zotero preservation is not an obligation.
  • Framework by study type (Ch. 7): prediction → TRIPOD+AI/PROBAST+AI; imaging-AI model → CLAIM 2024; radiomics → CLEAR; feature computation → IBSI 1/2.
  • When a paper moves to active reading, open it with the role/question above, not a blank read.
Adrian Tkacz, MD, MSc Computer ScienceAboutLinkedInORCIDXEmail

© 2026 Adrian Tkacz · BioMedical-IT
Original text and figures are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0), unless otherwise noted. Please credit Adrian Tkacz / BioMedical-IT and link to the original source. Third-party materials retain their respective rights and licenses.