Machine Learning Workflows in Radiomics

A radiomics study is a pipeline, and every stage of it can invalidate the stages after it. Read left to right; a weakness early is not recoverable later.

The five stages

1. Acquisition. Images arrive from a scanner under a protocol. Slice thickness, reconstruction kernel, dose and vendor all change the voxel values every later step depends on.
2. Segmentation. A region of interest is delineated, manually or automatically. Shape and boundary features move with contour variability.
3. Feature extraction. Hundreds to thousands of descriptors are computed. Preprocessing choices — resampling, grey-level discretisation, normalisation — are part of the feature definition, not a detail beneath it.
4. Feature selection and modelling. The candidate set is reduced and a model is fitted. This is where high-dimensional screening on a small cohort manufactures findings.
5. Validation. Internal, then external. Without the external step the pipeline has described one institution, not a biomarker.

Where pipelines fail

• Feature stability never tested, so unstable descriptors survive into the model.
• Data split made by image or lesion rather than by patient, letting one patient sit on both sides.
• Preprocessing under-reported, making the work impossible to reproduce and therefore impossible to refute.
• No comparison against established clinical predictors, so added value is unknown.