Synthetic Imaging and Invented Information

The source-to-target information gap is the central question

In the source-conditioned cases considered here, synthetic imaging generates a target image or contrast from an acquired source. The important question is not which image-to-image architecture is fashionable. It is what the source actually measured, what it only correlates with, and what the target asks the model to estimate. Any target content that was not directly acquired is a conditional, model-derived estimate from measured signals and population correlates, not a verified new patient measurement. The distinction is continuous rather than a binary law: a source can contain direct evidence, degraded evidence, anatomical correlates, or no reliable patient-specific evidence for a target property.

Low-dose to standard-dose reconstruction may be framed as estimating a cleaner representation from measurements in which relevant structure is degraded by noise. That framing does not guarantee fidelity. A learned operation can still suppress a subtle structure or add a pattern that was not supported by the acquisition. MR-to-synthetic-CT asks for an attenuation-related target from MR signals that do not directly measure CT attenuation; anatomical and protocol correlates may help, but the result remains a conditional inference. Translation between MR contrasts, protocol harmonisation, and non-contrast to contrast- enhanced CT each have their own information assumptions and clinical consequences. The risk of a misleading estimate varies continuously with the task and target property; it is not a binary distinction between denoising and synthesis.

The amount of target information that is absent or only indirectly represented is a useful risk question, not a universal risk score. Enhancement in a non-contrast image is not directly measured just because the anatomy may contain correlates of where enhancement often occurs. Conversely, denoising a measured signal does not guarantee that every clinically relevant detail survives. A synthetic image must therefore be treated as a conditional model output from measured signals and correlates, not as an acquired measurement or verified new patient information by construction.

Realistic appearance is not evidence of correctness

An output can look convincing while being anatomically inconsistent, changing a quantitative value, dropping a real abnormality, or adding a plausible but false one. Pixel similarity and visual realism answer whether an output resembles a particular reference under a particular comparison. They do not by themselves answer whether the relevant anatomy and pathology were preserved, whether a measurement remains valid, or whether a downstream task benefits.

This is especially important when the target was not acquired for the patient. A model can use population regularities to produce a likely-looking target, but population plausibility is not patient-specific confirmation. The more a claim depends on a property that was not directly measured, the more the evaluation must test that property through an independent reference, a defined task, or a carefully bounded interpretation.

Evaluation must follow the intended use

The evaluation should begin by naming the target and the use. Anatomical consistency asks whether structures and their relationships remain plausible. Pathology preservation asks whether abnormalities are retained rather than removed or fabricated. Quantitative evaluation asks whether declared values or features agree within a meaningful tolerance under the stated measurement definition. A downstream task or reader evaluation asks whether the synthetic image supports the decision it was intended to support. Uncertainty or an abstention mechanism may be relevant when the model is being asked to infer poorly observed information. External or otherwise independent testing asks whether these results travel beyond the data used to develop the method.

These are separate evaluation dimensions, not a universal checklist. Select the required tests from the intended use and the risks of a wrong output. Use a reader assessment only when readers are part of the claim being made.

No one of those checks substitutes for the others. A synthetic image can pass a visual comparison and fail a lesion-detection task, or preserve a global intensity summary while changing a boundary-sensitive feature. The appropriate set of checks depends on the claim; a paper that reports only a generic image similarity score has not thereby answered a clinical or quantitative-use claim.

An evaluation chain separates the claims

A useful design links separate evidence rather than treating one score as a proxy for the whole chain. Depending on the intended use and risk, a study may separate development from independent image evaluation, add quantitative and pathology checks, and test the intended use through a downstream task. A reader assessment belongs only when readers are part of the claim. The target reference and the source of independence must be explicit. A paired real target, a held-out subject, and an external site answer different questions, so the result should say which one was used.

The same chain exposes the limitation of a visually strong but unacquired target. If a model produces synthetic enhancement, the evaluation must address whether the enhancement is patient-specific and whether it changes the task decision, not merely whether it resembles contrast-enhanced images in aggregate. If a model denoises, the evaluation must still examine subtle structures and quantitative consequences rather than assuming that lower apparent noise is the objective. The model class does not choose the evaluation; the intended use does.

Dated examples belong beside the stable question

Synthetic-imaging papers, preprints, and model releases are useful for showing how a particular study operationalised these checks, but their identities, versions, and conclusions require re-verification as the field changes. The maintained companion surface carries the current reading list and model-specific details. This chapter keeps the durable distinction between measured, degraded, correlated, and inferred information instead of turning a current example into timeless doctrine. An unresolved presentation or webinar is not a reason to invent a paper identity or a result.

From invented information to model evaluation

The central discipline is to state what the source contains, what the model is asked to infer, and which independent checks can expose a wrong inference. Synthetic images may be useful conditional representations or estimates for a bounded task, but they do not automatically add verified information to the patient’s record. The next section applies this task-first evaluation logic to models more broadly, then asks how a performance difference between subgroups should be investigated without turning the first gap into a verdict.