Skip to content
Imaging Evidence Path

Why binWidth changes the answer

Tonight · ~20 min · read + a tiny calculation · energy: low · setup: none

Change one number in your PyRadiomics config — binWidth, from 25 to 5 — and a texture feature can move by a large fraction, on the same tumour. This lesson shows why with a five-value example, so you never again treat discretisation as a boring default.

The anchor: discretisation is the texture feature

Recall from lesson 3: texture matrices are built on the discretised image. In PyRadiomics’ fixed-bin-width rule each voxel’s grey level is floor(HU / binWidth), then shifted by a constant so the ROI’s minimum becomes 1 — a relabeling that does not change which voxels share a bin. So for the question “how many distinct grey levels exist, and which HU values collapse together?”, the floor(HU / binWidth) part is what decides it. The bin width decides how many distinct grey levels exist, which decides the entire matrix, which decides every feature. So binWidth is not a numerical-precision knob — it is part of the definition of the feature.

A five-value demonstration

Take five HU values inside a small ROI — three lung-adjacent and two soft-tissue:

HU values:  [ -987,  -982,  -978,   -50,   -30 ]

Apply floor(HU / binWidth) for two choices of bin width:

HU binWidth = 5 binWidth = 25
−987 floor(−987/5) = −198 floor(−987/25) = −40
−982 floor(−982/5) = −197 floor(−982/25) = −40
−978 floor(−978/5) = −196 floor(−978/25) = −40
−50 floor(−50/5) = −10 floor(−50/25) = −2
−30 floor(−30/5) = −6 floor(−30/25) = −2

Look at the lung triplet (−987, −982, −978):

  • At binWidth = 5 they become three different grey levels (−198, −197, −196) — the fine HU differences survive as distinct bins, and a GLCM would see three co-occurring levels.
  • At binWidth = 25 they collapse into one grey level (−40) — all three HU values are now “the same”, and the GLCM sees a single level where there were three.

The whole lung region, which at binWidth = 5 had internal texture, becomes a uniform blob at binWidth = 25. Every texture feature — Joint Entropy, Contrast, run lengths, zone sizes — is now computed on a different matrix and gives a different number, on identical anatomy.

Which way do features move?

The direction is usually predictable, which makes the effect auditable rather than mysterious:

  • Smaller binWidth → more grey levels → a finer, more populated matrix → typically higher entropy (more disorder visible) and finer-grained contrast.
  • Larger binWidth → fewer grey levels → a coarser, sparser matrix → typically lower entropy and coarser contrast.

(PyRadiomics’ default binWidth is 25; do not leave it implicit — pin it in the parameter file and report it with every result, as TRACE-CT does, because the chosen value is feature-determining.)

The discipline this forces

Because binWidth rebuilds the matrix, two statements follow:

  1. A feature value is meaningless without its binWidth (and resampling spacing, and interpolator). “JointEntropy = 2.3” tells you nothing; “JointEntropy = 2.3 at binWidth = 25, 1 mm isotropic B-spline, HU clip [-1000, 400]” is a quantity.
  2. You must not tune binWidth to maximise a downstream metric — that is feature-fishing and invalidates validation. You pin it, report it, and keep it fixed.

This is why IBSI fixes a processing workflow and why TRACE-CT commits configs/pyradiomics_params.yaml (binWidth 25, original, 3D, label 1) rather than leaving defaults. The configuration is part of the measurement.

Bridge

The five-value table is the whole argument compressed: the same five voxels produce either a textured or a uniform region depending on a single number in a config file. When you internalise that, “report your preprocessing” stops being bureaucracy and becomes obviously necessary — without binWidth, the number is not even defined.

Stop and think — then reveal

You compute original_glcm_JointEntropy at binWidth = 25, then again at binWidth = 5, on the same tumour. The value rises from 2.3 to 3.1. Has the tumour become “more textured”?

No — the tumour is unchanged; you changed the definition of the feature. Smaller binWidth preserves finer grey-level distinctions, populating more matrix cells and raising entropy by construction. The 2.3 and the 3.1 are two different quantities under two different pipelines, not a measurement of a change in the tumour. This is exactly why comparing a feature across papers (or across your own experiments) is invalid unless binWidth and the rest of the pipeline match.

What to retain

  1. binWidth discretises HU into grey levels; it rebuilds the texture matrix, so it changes every texture feature on identical anatomy.
  2. Smaller binWidth → more levels → higher entropy/finer contrast; larger → coarser → lower. The direction is usually predictable.
  3. A feature value is undefined without its binWidth (and spacing/interpolator). Pin it in the config, report it, never tune it to a metric.
  4. This generalises: discretisation is part of the feature’s definition, which is why IBSI and TRACE-CT fix it rather than leave it to defaults.

Next: zoom out — everything that changes a feature, of which binWidth is only the most dramatic.