Skip to content
Imaging Evidence Path

When synthetic/multimodal MRI asks the model to invent information

Tonight · ~20 min · read · energy: medium · setup: none

This lesson ties MRI back to Chapter 6’s synthetic-imaging question: for each MRI synthesis task, what information is actually present, and what is the model being asked to invent? Answer that, and the hallucination risk of each task becomes predictable.

The anchor: present vs invented, on concrete MRI tasks

Chapter 6 separated “denoising present information” from “inventing absent information”. MRI gives the cleanest examples, because MRI’s contrasts are partly independent measurements of the same anatomy:

  • T2 → T1 (or T2 → STIR) sequence synthesis — plausible. The same anatomy is measured; the model transforms one contrast into another. The information is largely present (same protons, same anatomy), so this is closer to translation than invention. Still verify it: a plausible-looking T1 is not guaranteed to be quantitatively faithful.
  • MRI → synthetic CT — the model infers electron density MRI cannot directly see. This is needed for MR-based radiation-therapy dose planning. The anatomy is present; the electron density (what CT measures) is not directly measured by MRI, so the model infers it from MR-visible cues. More invention than T2→T1.
  • Pre-contrast → post-contrast T1 — the model invents enhancement that was never measured. Gadolinium enhancement depends on perfusion and vascular permeability that the pre-contrast scan does not contain (lesson 3). This is the hardest case — the target is genuinely absent from the source.

The information-invention problem, restated for MRI

Pre-contrast MRI does not directly observe gadolinium enhancement — enhancement depends on perfusion, vascular permeability, and tissue biology the pre-contrast scan does not measure. So a model that synthesises post-contrast from pre-contrast is being asked to infer/invent contrast uptake from indirect cues. This is a fundamentally different problem from denoising or sequence translation (Ch. 6; EuSoMII notes §5).

Same architectures, fundamentally different problems — and the hallucination risk scales with how much the target is genuinely absent.

Why this matters for how you read a synthetic-MRI paper

A synthetic-MRI evaluation that reports only pixel realism (SSIM/PSNR) is insufficient for all of these, but especially for the invention-heavy ones. Demand (Ch. 6):

  • anatomical consistency — plausible-but-wrong structure;
  • pathology preservation — does disease survive, neither lost nor fabricated?;
  • quantitative / radiomic stability — do measurements agree on synthetic vs real?;
  • task utility — does a downstream read/model perform as well on synthetic?;
  • uncertainty — does the model know when it is guessing?;
  • hallucination risk — highest for pre→post-contrast, lower for T2→T1.

Connection to EuSoMII and TRACE-CT

The 2026-08-10 EuSoMII webinar framed this whole thread (see notes/2026-08-10-eusomii-msk-ai.md §5). TRACE-CT has a future missing-acquisition generation direction; this lesson is the conceptual scaffolding for reading that literature intelligently. PLAUSIBLE INFERENCE that synthesis will matter clinically; PROJECT HYPOTHESIS that the missing-acquisition framing is the right TRACE-CT use case. (The specific webinar T2→T1/STIR example could not be reliably identified and is preserved as explicitly unresolved in the literature map.)

Stop and think — then reveal

Rank these MRI synthesis tasks by hallucination risk (lowest to highest) and justify the order: (a) T2→T1 sequence translation, (b) low-dose-style denoising of a T2, (c) MRI→synthetic-CT, (d) pre-contrast→post-contrast T1.

Lowest to highest: (b) < (a) < (c) < (d). (b) denoising removes degradation from information that is present — least invention. (a) T2→T1 translates between contrasts of the same measured anatomy — mostly present, some contrast-specific invention. (c) MRI→CT asks the model to infer electron density MRI does not directly measure — more invention, but anatomy is present. (d) pre→post-contrast invents enhancement (perfusion/ permeability) entirely absent from the source — the most invention, the highest hallucination risk. The order tracks how much of the target is genuinely present in the source — the single question from Chapter 6.

What to retain

  1. Rank MRI synthesis tasks by what is present vs invented: denoising < contrast translation (T2→T1) < MRI→CT < pre→post-contrast. Hallucination risk scales with absence.
  2. Pre→post-contrast synthesis asks the model to invent enhancement never measured — fundamentally different from denoising/translation.
  3. Evaluate beyond pixel realism: anatomy, pathology, quantitative stability, task utility, uncertainty, hallucination risk.
  4. TRACE-CT’s future missing-acquisition direction uses this scaffolding; treat the clinical-impact claim as plausible inference, not established evidence.

You have finished Chapter 8. The sequence-family table, modality orientation, and further reading are in the MRI reference. Continue to Chapter 9 — Clinical deployment.