Radiomics

In short: From image acquisition to a validated model: each step adds variability and each needs reporting.

Steps

• Acquisition and reconstruction: scanner, protocol, slice thickness, kernel.
• Segmentation: drawing the region of interest by hand, semi-automatically or with AI.
• Pre-processing: resampling to common voxel sizes, intensity discretisation, filtering.
• Feature extraction: shape, first-order, texture and filtered features.
• Feature selection and modelling: reducing hundreds of features and fitting a model.
• Validation: internal resampling, then independent external data.

Reporting matters

Two papers can compute the "same" feature and get different numbers because of a pre-processing choice. Reporting checklists such as the Radiomics Quality Score and the CLEAR checklist help readers judge what was done.

In short: What each family measures and why texture features need extra care.

Shape

Volume, surface area, sphericity and elongation describe geometry. They depend on segmentation more than on intensity.

First-order

Mean, variance, skewness, kurtosis and entropy summarise the distribution of intensities without regard to where voxels are.

Texture

Matrices such as the grey-level co-occurrence matrix (GLCM) count how intensities relate to their neighbours, giving features like contrast and homogeneity. They depend heavily on discretisation and voxel size.

Higher-order and learned

Applying wavelet or Laplacian-of-Gaussian filters before extraction multiplies the number of features. Deep learning can learn its own representations, which can be powerful but are harder to interpret and to standardise.

In short: Test-retest scans, inter-scanner variation, segmentation variability and harmonisation, in plain language.

Repeat the scan, get a different number

Same-day repeat scans of the same patient, such as the public RIDER lung CT collection, show that some features barely change while others swing widely. Features that are not repeatable cannot be good biomarkers.

Scanner and protocol effects

Different manufacturers, kernels and slice thicknesses shift feature values. Statistical harmonisation methods such as ComBat, borrowed from genomics batch correction, can reduce site effects but can also remove real signal if misapplied.

Segmentation variability

Two readers outlining the same lesion produce different masks; texture and shape features inherit that disagreement. Studies often report intraclass correlation coefficients to show stability.

In short: AI models can segment and classify images, but performance depends on data, and research code is not a regulated medical device.

Ground truth is a choice

AI segmentation models learn from human-drawn labels. Those labels carry reader variability and sometimes errors, so "ground truth" is a reference standard, not truth.

Domain shift

A model trained on one scanner, population or protocol may perform worse elsewhere. External validation on data from new sites is essential, and performance can drift as practice changes.

Research code versus regulated software

Open-source frameworks such as MONAI or nnU-Net are research tools. Software intended for diagnosis or treatment decisions is regulated as a medical device in many jurisdictions. RadioGx labels repository results as community developer references.