Skip to content
Imaging Evidence Path

Chapter 3 — Quantitative imaging and radiomics

This chapter · ~3 h total across 8 lessons · The chapter where the pipeline’s choices become numbers · Energy: low to active · Setup: Python (pyradiomics, SimpleITK) + RIDER data for the hands-on lesson

Radiomics promises to turn an image into a table of features you can model. The promise is real; the failure modes are legendary. The field has a reproducibility problem precisely because the pipeline is long and almost every step is a choice that silently changes the numbers. This chapter does not teach radiomics as a feature glossary — it teaches the few ideas that decide whether a number means anything.

The thread this chapter pulls

A radiomic feature is not a property of a tumour. It is a property of a tumour under a stated pipeline: a particular image, mask, geometry and preprocessing. Change any of those and the number changes, even though the biology did not. So the chapter is built around two questions: how does an image become a table of features, and can a given feature be trusted?

  • how a CT plus a mask becomes hundreds of candidate measurements, and why most of them are correlated and unstable;
  • how texture is actually turned into numbers — discretisation first, then the matrices — with a worked example, not an acronym list;
  • why binWidth changes the answer, with a tiny numerical demonstration;
  • what changes a feature (spacing, reconstruction, segmentation, preprocessing) and why each is a variable you must report;
  • repeatability vs reproducibility, ICC/CCC, and test–retest on RIDER;
  • harmonisation and ComBat’s limits, robustness filtering, and the handoff to a model.

Lessons

  1. How an image becomes a table of measurements — quantitative imaging → the radiomics pipeline.
  2. One tumour, hundreds of candidate features — first-order, shape, texture intuition.
  3. How texture is turned into numbers — discretisation first, then a worked GLCM, then the matrix families.
  4. Why binWidth changes the answer — a tiny numerical example.
  5. The same tumour can produce different radiomics — spacing, reconstruction, segmentation, preprocessing.
  6. Can this feature be trusted? — repeatability vs reproducibility, ICC/CCC, test–retest.
  7. RIDER as a robustness experiment — a real practical workflow.
  8. From feature to evidence — harmonisation, ComBat’s limits, robustness filtering, the model handoff.

The full feature-family reference table (GLCM/GLRLM/GLSZM/GLDM/NGTDM, filtered features, the IBSI/QIBA/CLEAR standardisation stack, the tools) lives in the radiomics reference.

How this chapter connects

  • It assumes Chapter 1 (working in physical HU) and Chapter 2 (a feature is a function of image, mask, geometry).
  • Chapter 4 takes the feature table and asks whether it supports an honest model — leakage, p >> n, calibration.

If you only have an hour, lessons 3 (texture into numbers) and 4 (binWidth) are the core; lesson 6 is the robustness idea the rest of the field runs on.