Skip to content
Imaging Evidence Path

How an image becomes a table of measurements

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

A CT and a tumour mask go in; a spreadsheet of hundreds of numbers comes out. That transformation is the radiomics pipeline, and it is not magic — it is a chain of choices, every one of which changes the numbers. This lesson is the map of that chain, so that when a feature looks surprising you know which link to interrogate.

The anchor: three terms that are not synonyms

Before the pipeline, disentangle three words people use loosely:

  • Quantitative imaging — extracting objective, reproducible measurements from images (volumes, diameters, attenuation statistics). The broadest, most defensible term. ESTABLISHED EVIDENCE that this is feasible and useful.
  • Radiomics — a subset: extracting a large battery of (mostly texture and shape) features, then modelling them statistically. Higher-dimensional, higher reproducibility risk.
  • Imaging biomarker — a feature validated against a biological/clinical reference (e.g. predicts survival independently of known confounders). The strongest claim. Most radiomic features are not biomarkers. Calling a texture feature a “biomarker” without validation is an overclaim.

These three sit inside a broader measurement framework — QIBA — which characterises the bias and repeatability of quantitative measurements themselves (reference).

The pipeline, read as a chain of choices

flowchart LR
    A["acquisition /<br/>reconstruction"] --> B["segmentation<br/>(ROI)"]
    B --> C["preprocessing"]
    C --> D["resampling"]
    D --> E["intensity<br/>discretisation"]
    E --> F["feature<br/>extraction"]
    F --> G["robustness<br/>analysis"]
    G --> H["model /<br/>interpretation"]

Every arrow is a step that changes the features. The central discipline is to pin each step and report it. IBSI exists largely to standardise these pins.

  • Step 1 — acquisition / reconstruction. Record kernel, slice thickness, kVp, dose, contrast phase, scanner, site. Two reconstructions of one scan are different images (Ch. 1). These become covariates or stratification variables.
  • Step 2 — segmentation (the ROI). The ROI defines what you measure. A one-voxel boundary error changes shape and edge-sensitive texture features. The ROI’s reference geometry must match the image (Ch. 2). Whether manual, semi-automatic, or a model output (e.g. TotalSegmentator) is itself a variable.
  • Step 3 — preprocessing. Filtering, intensity windowing, grey-level normalisation. Keep preprocessing deterministic and parameter-fixed (no cohort-level statistics unless leakage is handled — Ch. 4). TRACE-CT clips HU to [-1000, +400] before extraction.
  • Step 4 — resampling. Normalise to a common spacing so texture features are comparable across scans. B-spline for the image, nearest-neighbour for the mask; report the target spacing (Ch. 2).
  • Step 5 — intensity discretisation. Most texture features are not computed on raw HU. The intensity range is quantised into grey levels via a fixed bin width. This step is so consequential it gets its own lesson.
  • Step 6 — feature extraction. Run the feature definitions on the (preprocessed, resampled, discretised) image + mask (lesson 2).
  • Step 7 — robustness analysis. Is the feature stable? Test–retest, perturbation, reconstruction/scanner dependence (lesson 6). Filter by robustness before modelling, inside CV.
  • Step 8 — model / interpretation. Only now fit a model. The reproducibility problem is most visible here (hundreds of features, small cohorts, optimistic reporting) but the failures are usually decided upstream at steps 4–7.

The discipline this map forces

A feature “changed” between two studies is almost never a biological discovery; it is a difference somewhere in this chain — a different kernel, a different binWidth, a different mask. The pipeline map is the diagnostic tool: when a number is surprising, walk the chain and ask which pin differs. That is why TRACE-CT commits a parameter file (configs/pyradiomics_params.yaml) with fixed binWidth and spacing, and records the reconstruction role of every series — so the chain is auditable, not implicit.

Stop and think — then reveal

Two papers report “GLCM JointEntropy” for the same cancer type and get very different distributions. Neither is necessarily wrong. What four pipeline pins must agree before you can attribute the difference to biology?

  1. Reconstruction (kernel, slice thickness — step 1).
  2. Resampling target spacing and interpolator (step 4).
  3. Discretisation — binWidth and the binning method (step 5).
  4. Segmentation — how the ROI was defined and on what geometry (step 2).

If any of these differ, the two “JointEntropy” values are different quantities under different pipelines, not comparable measurements. You cannot read a biology-vs-biology difference off them until the pins match.

What to retain

  1. Quantitative imaging ⊃ radiomics; “biomarker” is a validated claim most radiomic features have not earned.
  2. The pipeline is a chain of choices (acquisition → segmentation → preprocessing → resampling → discretisation → extraction → robustness → model); each changes the features.
  3. The discipline is to pin and report every step. IBSI standardises the pins.
  4. Surprising features are usually decided upstream — walk the chain to find which pin differs before invoking biology.

Next: what actually comes out of step 6 — the feature families and their intuition, before the matrices in detail.