Skip to content
Imaging Evidence Path

Reconstruction can change the biology you measure

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

A model trained on filtered back-projection images may not behave the same on iterative- or DL-reconstructed images, and radiomic features computed on DL-reconstructed data are not guaranteed to equal those on FBP. This lesson is the reconstruction→quantification link, and why the reconstruction method is a variable that changes the biology you think you see.

The anchor: three reconstruction regimes, three textures

CT reconstruction has a conceptual progression, and each regime produces a different texture from the same raw data:

flowchart LR
    F["FBP<br/>filtered back-projection"] --> I["Iterative Reconstruction<br/>(IR)"]
    I --> D["DL Reconstruction"]
  • FBP — the classical analytic reconstructor; deterministic, well-understood, noisy at low dose.
  • Iterative reconstruction (IR) — models noise/physics/statistics; lower noise at a cost of altered texture (the “plastic” appearance) — which changes texture features.
  • DL reconstruction — a learned mapping (often from low-dose/raw to standard-dose-like); can reduce dose or speed acquisition, but can also alter or invent texture, changing radiomic features and potentially suppressing or appearing pathology.

Why this matters for quantitative imaging: a model trained on FBP images may not behave the same on IR or DL-reconstructed images, and radiomic features computed on DL-reconstructed data are not guaranteed to equal those on FBP. Always record the reconstruction method (Ch. 1, Ch. 3).

Why this is a quantitative problem, not a cosmetic one

Reconstruction is part of the feature-defining pipeline (Ch. 3): the feature is a function of (image, mask, geometry, preprocessing), and the reconstruction method shapes the image. So two cohorts reconstructed differently are not directly comparable, and a model trained on one regime is being tested under domain shift when deployed on another (Ch. 4). DL reconstruction adds a sharper risk: by learning the image, it can remove noise that carried information or fabricate texture that looks plausible — the same “what is the model inventing?” question that defines synthetic imaging (lesson 6).

The practical discipline

  • Record the reconstruction method with every study (FBP / IR vendor+level / DL model+version), as you would record kernel and dose.
  • Do not pool across reconstruction methods without harmonisation and intent (Ch. 3).
  • Test under the reconstruction shift you expect at deployment — a model that must read IR-reconstructed images should be tested on them, not only on FBP.
  • Be suspicious of DL reconstruction changing a quantitative result in a flattering direction; verify the change is biology, not the reconstructor inventing texture.

This is exactly why TRACE-CT tags the reconstruction role of every series (classify_source_role(): lung_1_25mm vs standard_5_0mm) — the reconstruction is a first-class variable, not a metadata afterthought.

Stop and think — then reveal

A radiomics study trains on FBP-reconstructed CTs and reports a texture biomarker. The deployment site has since switched to an iterative-reconstruction default. Is the biomarker still valid there, and what would you demand before trusting it?

Not automatically. Iterative reconstruction changes the texture distribution (“plastic” appearance, altered noise), so the features the biomarker depends on are computed on a different effective image — a reconstruction-shift domain change (Ch. 4). Before trusting it you would demand frozen-model evaluation on IR-reconstructed CTs from the deployment site (and recalibration if the distribution shifted). A biomarker “validated” only on FBP is licensed for FBP, not for the new regime.

What to retain

  1. FBP → IR → DL reconstruction each produce a different texture from the same raw data; IR and DL change texture features.
  2. Reconstruction is part of the feature-defining pipeline, so cross-reconstruction pooling is a domain shift; models may not transfer across regimes.
  3. DL reconstruction can invent or suppress texture — the “what is being invented?” question bridges to synthetic imaging.
  4. Record the reconstruction method, do not pool across methods, test under the deployment regime, and verify flattering quantitative changes are biology.

Next: the extreme case of “what is the model inventing?” — synthetic imaging.