Skip to content
Imaging Evidence Path

Note — EuSoMII 2026: MSK and generalist imaging AI (2026-08-10)

A learning note, not minutes. It records the conceptual models I took away from the lectures, with claim-status labels and pointers into the literature map. Citations are resolved in resources/literature-map.md (PMIDs/DOIs there); PDFs live in Zotero. Status vocabulary: ESTABLISHED EVIDENCE · EXPERT CONSENSUS · COMMON PRACTICE · PLAUSIBLE INFERENCE · PROJECT HYPOTHESIS.

1. Quantitative radiology — three terms that are not synonyms

The dominant framing of the meeting: radiology is becoming quantitative. But three terms were used loosely and must be kept distinct:

  • Quantitative imaging — extracting objective, reproducible measurements from images (a volume, a diameter, an attenuation statistic). The broadest, most defensible term. ESTABLISHED EVIDENCE that this is feasible and useful.
  • Radiomics — mining a large battery of (largely texture) features and modelling them. Higher-dimensional, higher-reproducibility-risk (see learning/03). EXPERT CONSENSUS that it is promising; reproducibility is the open problem.
  • Imaging biomarker — a measurement validated against a biological/clinical reference. The strongest claim; most radiomic features are not biomarkers. COMMON PRACTICE to overuse the term; rigorously it requires validation.

Why I retain this: the confusion is load-bearing. A paper that reports a texture feature and calls it a biomarker has overclaimed; a measurement that is not robust to reconstruction is not even a stable quantitative-imaging value.

2. Segmentation as infrastructure, not a destination

The interesting clinical path is rarely image → diagnosis. It is:

image → segmentation → measurement → quantitative phenotype
      → prediction / diagnosis / longitudinal assessment

Segmentation is infrastructure: the thing that turns a picture into a measurable object. Two concrete examples from the meeting/literature:

  • TotalSegmentator (Wasserthal et al., Radiol AI 2023; PMID 37795137, VERIFIED) — segmentation of ~104 structures in one pass, enabling downstream body-composition and organ measurements. The model-description is less interesting than its role: a reusable mask factory.
  • Quantitative body composition (Ziegelmayer et al., Lancet Reg Health Eur 2025; PMID 40487774, VERIFIED) — IMAT/lean-muscle measurement in chronic back pain. This is the segmentation → measurement → phenotype → clinical question chain done on real patients — exactly the spine of this learning path and of TRACE-CT’s segmentation→radiomics pipeline.

Why I retain this: it reframes my learning priority. Robust, geometry-correct segmentation (Chapters 2–3) is not a side skill; it is the bottleneck on the whole quantitative chain.

3. The AI-assisted clinical workflow (the pipeline to hold in mind)

DICOM ──┬─ prior imaging
        ├─ clinical data
        └─ quantitative processing / segmentation
              → multimodal model
              → assisted interpretation / structured reporting
              → radiologist (final authority)

EXPERT CONSENSUS on the shape; PLAUSIBLE INFERENCE on how much of it is deployed reliably today. Three things I took away:

  • Multimodal is the direction: imaging plus clinical/demographic/prior data, not image-only models (this is also the statistically honest framing in learning/04 — does imaging add value beyond the clinical baseline?).
  • Assisted, not autonomous: the radiologist remains the authority. This is a clinical, medico-legal, and workflow reality, not just caution.
  • The pipeline is where failure modes live — identity/provenance loss, mismatched priors, segmentation silently wrong, AI result not retrievable in PACS. These are imaging-informatics problems (learning/05), and they matter as much as model accuracy.

4. Foundation / generalist models — a direction, not tonight’s task

Recorded as an important horizon. Statuses below updated to the verified literature map (resources/literature-map.md); the webinar narrative is unchanged.

  • Merlin — CT vision-language foundation model, Nature 2026, PMID 41781626 (earlier preprint PMID 38978576). Peer-reviewed (VERIFIED); fast-moving.
  • a2z-1 (Rajpurkar et al., arXiv:2412.12629, VERIFIED) — abdomen-pelvis CT generalist with external validation across many conditions. Preprint, not yet peer-reviewed. Strong external-validation design — useful as a methodology reference even before final publication.
  • Foundation-model reviews — Akinci D’Antonoli et al., Diagn Interv Radiol 2026;32(3):259-272, PMID 40626693 (VERIFIED peer-reviewed review); and Paschali et al., Radiology 2025;314(2):e240597, PMID 39903075 (VERIFIED). Useful orientation.
  • OmniMRI — unified VLM foundation model for generalist MRI, He et al., arXiv:2508.17524 (VERIFIED preprint, 2025-08-24). Fast-moving.
  • Agentic / MCP orchestration — ReCo (Orthanc-centred, TotalSegmentator/Merlin) — Tzanis & Klontzas, medRxiv 2026.07.14.26358025 (VERIFIED preprint). Speculative without strong clinical validation yet.

Why I retain this: generalist/foundation models are the likely medium-term substrate for both segmentation infrastructure and assisted interpretation. But they are PLAUSIBLE INFERENCE as to clinical impact today, and several key items are preprints. They are covered readably in learning/06-medical-imaging-ai — foundation models; read for orientation and informed paper reading, not immediate implementation. (Their detailed clinical validation remains a research frontier.)

5. Synthetic imaging — the most TRACE-CT-relevant concept

The meeting surfaced several “synthetic image” problems. They share image-to-image / conditional-generation methods but are not equivalent:

Problem Source → target What must be invented
MRI → synthetic CT MR anatomy → CT attenuation electron density MR doesn’t show
T2 → synthetic T1/STIR one MR contrast → another a contrast sequence not acquired
low-dose → standard-dose CT noisy CT → clean CT noise removal (information present, degraded)
reconstruction harmonisation one kernel/protocol → another a transform between protocols
non-contrast → contrast-enhanced CT pre-contrast → post-contrast contrast uptake (genuinely absent)

The key question (the most important thing I took from this thread):

What information is actually present in the source image, and what is the model being asked to invent?

Low-dose → standard-dose mostly denoises information that is present. Non-contrast → contrast asks the model to hallucinate contrast enhancement that was never measured. These are fundamentally different physical and clinical problems, even though both may use a pix2pix/cycle/diffusion architecture.

Why realistic appearance is insufficient. A synthetic image can look photorealistic and still:

  • violate anatomical consistency (a plausible-looking but wrong structure);
  • drop or invent pathology (the clinically fatal failure);
  • change radiomic/quantitative values in ways invisible to the eye;
  • be useless or misleading for the downstream task.

So evaluation must go beyond pixel realism (SSIM/PSNR/visual realism) to:

  • anatomical consistency
  • pathology preservation
  • quantitative / radiomic stability
  • task utility (does a downstream model/read perform as well on synthetic as real?)
  • uncertainty / calibrated confidence
  • external testing (not just in-distribution)
  • hallucination / invented-information risk

Relation to TRACE-CT: TRACE-CT has an explicit future missing-CT-acquisition generation direction (conditional synthesis to fill a missing acquisition/sequence). The current pre-Spark experiment is intentionally bounded and is not modified by this note — per the project’s anti-drift rule. When that direction activates, the evaluation dimensions above become the checklist; until then this is durable conceptual preparation (recorded as the deferred C11 synthesis facets, D-011), not active work.

PLAUSIBLE INFERENCE that synthesis will matter clinically; PROJECT HYPOTHESIS that the missing-acquisition framing is the right TRACE-CT use case.

What I take into next week’s learning

  1. Keep quantitative imaging / radiomics / biomarker rigorously separate.
  2. Treat segmentation as infrastructure — it is the bottleneck; Chapters 2–3 are high priority.
  3. Hold the AI-assisted workflow pipeline as the context for imaging informatics (Ch. 5).
  4. Read foundation/generalist models via Ch. 6 Part F (orientation); watch the preprints via the literature map.
  5. For synthetic imaging, internalise “what is present vs what must be invented” and the beyond-pixel-realism evaluation list — it is the conceptual scaffolding for TRACE-CT’s future synthesis direction.