Skip to content
Imaging Evidence Path

The same tumour can produce different radiomics

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

binWidth is the most dramatic feature-determining variable, but it is far from the only one. This lesson is the full dependence map: every input to a feature that, if changed, changes the number. Hold this map and “why did my feature change?” becomes a finite search instead of a mystery.

The anchor: a feature is a function of four things

A radiomic feature is a function of (image, mask, geometry, preprocessing). We have already met three of the four:

  • image — the reconstruction (Ch. 1);
  • geometry — the resampling target and interpolator (Ch. 2);
  • preprocessing — the HU clip, normalisation, and discretisation (lesson 4).

The fourth is the mask — the segmentation itself. Change any of the four and the feature changes, even though the patient did not. So “what changes a feature?” has a finite, auditable answer.

The dependence map

Variable How it changes the feature Where controlled
Interpolation / resampling target spacing and interpolator change the matrix (Ch. 2) pipeline pin
Discretisation (binWidth) rebuilds the grey-level alphabet (lesson 4) pipeline pin
Segmentation dependence manual vs auto, boundary placement, ROI margin ROI definition
Voxel-size dependence many texture features scale with voxel size — why isotropic resampling is standardised resampling pin
Reconstruction / scanner / protocol kernel, slice thickness, dose, kVp, contrast phase, site acquisition record

Two of these deserve a sentence each, because they are easy to under-rate:

  • Voxel-size dependence. Many texture features are not invariant to voxel size: the same tumour scanned at 0.7 mm and 2 mm yields different matrices because the spatial relationships the matrix counts are sampled differently. This is why isotropic resampling to a fixed spacing is standard practice, and why the spacing must be reported — not as a nicety, but because it is a feature-determining variable.
  • Segmentation dependence. A one-voxel change in the ROI boundary moves voxels in and out of the region, changing first-order statistics and reshaping the texture matrices. Whether the mask was drawn by a human, grown by a tool, or produced by a model (TotalSegmentator) is itself a variable that propagates into every feature. This is the bridge to lesson 6: if a feature swings when you nudge the mask, it is not a biomarker.

The rule this map implies

There is one operational consequence, and it is absolute:

Pin every variable in this map; report it with every result; never pool across different settings without intending to.

This is why IBSI fixes a processing workflow, why PyRadiomics lets you pin resampling and binWidth, and why TRACE-CT commits a parameter file plus a reconstruction-role tag for every series. The feature is a property of the tumour under a stated pipeline; without the pipeline statement, the number is not even comparable to itself across runs.

Stop and think — then reveal

You read two radiomics studies of the same cancer type. Study A reports binWidth = 25, 1 mm isotropic, B70f lung kernel. Study B reports binWidth = 5, 2 mm isotropic, B40f soft kernel. Both claim their texture feature “distinguishes survivors”. Can you compare their features?

No — they are different quantities under different pipelines (different discretisation, different spacing, different kernel). Even the direction of a between-group difference could flip under a different binWidth, because the matrix is rebuilt. You can compare their conclusions only loosely, and only if both held out an external test set; you cannot compare the feature values themselves. This is the structural reason radiomics has a reproducibility problem, and the reason IBSI exists to standardise the pins.

What to retain

  1. A feature is a function of (image, mask, geometry, preprocessing). Change any of the four and the number changes.
  2. The dependence map: interpolation, binWidth, segmentation, voxel size, reconstruction/scanner/protocol. Each is a variable to pin and report.
  3. Voxel-size dependence is why isotropic resampling is standard and why spacing must be reported; segmentation dependence is why mask stability matters.
  4. The rule: pin every variable, report it, never pool across settings unintentionally. The feature is a property of the tumour under a stated pipeline.

Next: given that everything changes a feature, which features can you trust at all? — repeatability, reproducibility, ICC/CCC.