Skip to content
Imaging Evidence Path

From feature to evidence

Tonight · ~25 min · read · energy: medium · setup: none

You now have stable features. The last mile to a clinical claim is the hardest: harmonising across scanners, filtering robustly, and handing a feature table to a model without letting the modelling step reintroduce all the problems you just solved. This lesson is that last mile — and the boundary where radiomics hands off to Chapter 4.

The anchor: multi-centre features are not directly comparable

Scan the same tumour on a Siemens B70f and a GE soft kernel and the feature distributions differ — not because the tumour differs, but because the acquisition does (Ch. 1, lesson 5). Pool those features across sites into one model and the model can learn “which scanner” instead of “which tumour”. That is the multi-centre radiomics problem, and harmonisation is the attempt to remove the scanner/site effect.

ComBat — and its limits

ComBat adjusts for scanner/site batch effects on the feature distribution (originally from genomics). It is widely used to harmonise multi-centre radiomics. The limits you must know (EXPERT CONSENSUS on the limits; PLAUSIBLE INFERENCE on application):

  • ComBat assumes batch effects are independent of the biology/outcome. If a scanner difference is correlated with outcome (e.g. one site has both a different scanner and sicker patients), ComBat can remove real signal (over-correction) or leave residual bias.
  • It must be fit on training data only (inside CV), or it leaks — the test distribution shapes the harmonisation.
  • It harmonises distributions, not acquisition physics — it cannot recover information a different kernel destroyed. A texture wiped out by a smooth kernel is gone; ComBat cannot put it back.
  • There are variants (M-ComBat, B-ComBat, harmonisation of images vs features) that make different assumptions.

The honest framing: ComBat is a useful tool with assumptions that must hold; it is not a “make multi-centre radiomics work” button.

The robust filter, properly placed

The robustness filter from lesson 6 and any harmonisation both belong inside the cross-validation loop, in this order:

flowchart LR
    TR["training fold"] --> H["harmonise<br/>(ComBat on train only)"]
    H --> RF["robustness filter<br/>(stable on train only)"]
    RF --> FS["feature selection<br/>(on train only)"]
    FS --> M["fit model"]
    M --> TE["evaluate on<br/>held-out fold"]

Every box before “evaluate” uses only training data. Move any of them outside the loop and you have leaked test information into model development, producing optimistic performance — the central failure of Chapter 4. “Fit the scaler/ComBat/feature-selector on all patients before CV” is the single most common radiomics optimism source.

The model handoff

Only after stable, harmonised, leakage-safe filtering do you fit a model. At the handoff, three things are true and worth stating plainly:

  1. The feature table is now p features on n patients, almost always with p >> n (hundreds of correlated features, tens-to-hundreds of patients). Ordinary modelling intuition breaks here; regularisation and honest selection are mandatory (Ch. 4).
  2. The model must be benchmarked against a clinical tabular baseline (age, stage, histology) — the real question is whether imaging adds value beyond what the clinical record already provides, not whether the image model beats chance.
  3. A good internal AUROC is not a biomarker. Discrimination, calibration, external validation, and clinical utility must all hold before the feature table supports a clinical claim — the subject of Chapters 4 and 7.

The standardisation stack (where this all lands)

Radiomics reproducibility is a stack of layers, each owned by a different effort:

QIBA       characterises the measurement (bias/precision under a protocol)
   ↓
IBSI 1/2   standardise feature/filter computation + reference values
   ↓
CLEAR      reports the radiomics study (58 items, ESR/EuSoMII-endorsed)

QIBA defines “what a measurement is worth”; IBSI defines “how to compute the number”; CLEAR defines “what a study must report”. They are complementary layers of the same reproducibility stack (full citations in the reference).

Stop and think — then reveal

A multi-centre radiomics study harmonises features with ComBat fit on the entire dataset, then runs 5-fold cross-validation and reports AUROC 0.88. Why is the 0.88 not trustworthy as an honest performance estimate?

Because ComBat saw the whole dataset — including every test fold — before the CV splits were evaluated. The harmonisation (and thus the feature distributions the model trained on) was shaped by test-fold information: that is leakage, and it makes the reported AUROC optimistic. The fix is to fit ComBat inside each fold on training data only, then apply the fitted transform to the held-out fold. The same applies to scaling, feature selection, and the robustness filter: anything fit on data belongs inside the loop.

What to retain

  1. Multi-centre features are not directly comparable; ComBat harmonises distributions but assumes batch effects are independent of outcome, must be fit on training only, and cannot recover physics a different acquisition destroyed.
  2. Harmonisation, robustness filtering, scaling, and feature selection all belong inside CV — fit on training only, applied to the held-out fold.
  3. At the model handoff: expect p >> n; benchmark against a clinical tabular baseline; do not confuse a good internal AUROC with a biomarker.
  4. The reproducibility stack: QIBA (measurement value) → IBSI 1/2 (computation) → CLEAR (reporting). Different layers, complementary.

You have finished Chapter 3. For the full feature-family reference, the IBSI/QIBA/ CLEAR citations, the tools, and the failure-mode catalogue, see the radiomics reference. Then continue to Chapter 4 — Statistical learning and validation.