A radiologist looking at a tumour extracts a handful of descriptors: roughly how big, roughly what shape, margins smooth or spiculated, enhancement pattern, whether it has changed. Those descriptors are genuinely informative and they are also a drastic compression of what is in the image.
Radiomics is the proposition that the discarded information is worth recovering. Segment the lesion, run a battery of mathematical operators over the voxels inside it, and you get hundreds of numbers describing shape, intensity distribution and texture. Feed those numbers into a statistical model alongside clinical variables and see whether they predict something useful — response to a particular therapy, recurrence, survival.
The term was introduced in a 2012 European Journal of Cancer paper arguing that high-throughput extraction of quantitative features could turn routine scans into data for decision support, with the further hypothesis that imaging phenotype partly reflects underlying tumour biology. A 2016 Radiology review made the case more formally under the slogan that images are more than pictures, they are data.
Fourteen years on, the idea remains compelling and the field's track record is mixed in instructive ways. Both halves of that sentence are worth explaining.
What actually gets computed
The feature families are standard and there is nothing mysterious about them.
Shape features describe the segmented volume geometrically: volume, surface area, the ratio between them, how close to spherical it is, how convex, the ratio of its principal axes. These are the features closest to what a radiologist already assesses, and they are the most stable.
First-order features describe the histogram of intensities inside the region, ignoring where each voxel sits: mean, variance, skewness, kurtosis, percentiles, entropy of the intensity distribution. On a CT these are Hounsfield-unit statistics and carry real physical meaning. On most MRI sequences the intensities are not absolute, which matters enormously and I will come back to it.
Texture features are the interesting ones. They describe spatial relationships between voxel intensities, computed from matrices that count how often particular intensity pairs occur in particular geometric arrangements — co-occurrence, run-length, size-zone and neighbourhood difference matrices being the usual suspects. From these you derive things like contrast, correlation, homogeneity and various entropies. The biological story attached to them is that a heterogeneous texture reflects heterogeneous tissue — mixed cell populations, necrosis, irregular vascularity — and that heterogeneous tumours behave worse. The story is plausible and has some support. It is not established the way a histological grade is established.
Filtered features are all of the above recomputed after applying wavelet decompositions or Laplacian-of-Gaussian filters at several scales. This is where feature counts explode from dozens into the high hundreds or thousands, and where the statistical trouble begins.
The statistical trap, stated plainly
The defining methodological problem in radiomics is the ratio between features and patients.
A study with, say, 120 patients and 1,400 extracted features is not doing modelling in any comfortable sense. With that many candidate predictors, some will correlate with the outcome by chance alone, and a feature selection procedure run without proper nesting will find them reliably. The resulting signature looks impressive on the data it was selected on and evaporates on anyone else's.
This is not a hypothetical failure mode. It is the single most common reason radiomics signatures published with strong internal results do not replicate externally. A large fraction of the literature reports cross-validated performance where the feature selection happened outside the cross-validation loop, which is simply an error, and one that inflates apparent performance substantially.
If I am reviewing a radiomics study, the questions are always the same. How many features entered selection, and how many patients? Was selection nested inside the resampling, or run once on the whole dataset? Is there an external validation cohort from a different institution, and does the reported performance there hold up? Most of the time the answer to the last one is that there is no external cohort.
Reproducibility: the boring problem that decides everything
Even with clean statistics, a radiomic feature is only useful if it means the same thing twice.
Acquisition sensitivity. Texture features are computed from voxel intensities, so anything that changes voxel intensities changes them. Slice thickness, reconstruction kernel, tube current, field strength, sequence parameters, contrast timing. Two scanners imaging the same phantom will not produce the same texture values, and the differences are often larger than the biological differences the study is trying to detect. Test-retest studies — imaging the same patient twice within a short interval — consistently find that only a subset of features are stable, and that the unstable ones are disproportionately the exotic filtered ones.
Segmentation sensitivity. Most features are computed over a region someone drew. Two radiologists contouring the same lesion will not draw the same boundary, and shape and texture features respond to that. Features that survive inter-observer variation are a useful subset to restrict to, and restricting to them is a good discipline.
Preprocessing choices. Resampling to isotropic voxels, intensity discretisation bin width, normalisation scheme — all of these change the numbers, and for years they were reported inconsistently or not at all. Feature named the same thing in two papers were often not the same quantity.
The field's response has been standardisation. The Image Biomarker Standardisation Initiative produced consensus definitions and reference values for a large set of features, so that two implementations computing "GLCM contrast" now actually agree. Harmonisation methods borrowed from genomics can remove some scanner-related batch effects statistically. Neither eliminates the problem; both make it tractable. Anyone starting a radiomics project today who is not working against a standardised feature definition set is choosing to make their results uninterpretable.
Radiomics against deep learning
These are usually framed as rivals. In my experience the more useful framing is that they make different bets about where the prior knowledge lives.
Radiomics fixes the feature representation in advance, using operators chosen by humans, and then fits a relatively simple model on top. This means you need far less data, and the resulting model is interpretable in a specific and limited sense — you can say which features drove the prediction, though what a high wavelet-filtered run-length non-uniformity means physiologically is not obvious to anyone.
End-to-end deep learning learns its own representation, which is more powerful when there is enough data and less controllable when there is not. It also inherits the shortcut-learning problems that come with letting a network find whatever correlates with the label, including scanner identity.
For prognostic and predictive modelling on the cohort sizes typical of a single institution — a few hundred patients — the hand-engineered route is frequently the more honest choice. For detection and segmentation on large datasets, learned representations win clearly. There is also a growing middle ground of using learned features and engineered features together, which works better than the partisanship of a few years ago suggested it would.
Where this stands clinically
Radiomic signatures have shown genuine prognostic association across several tumour types, most extensively in lung, head and neck, and glioma. That is a real scientific result. The gap between that and a tool a clinician uses to choose a treatment is still substantial, and it is not primarily a modelling gap. It is prospective validation, acquisition standardisation, and integration into a workflow where someone has to segment the lesion before anything can be computed at all.
The most likely near-term path is not radiomics as a standalone product but radiomic features as one input among several — combined with genomics, pathology and clinical variables — in models whose value comes from the combination. That is less quotable than "the scan predicts survival," but it is closer to how the evidence is actually shaping up.
What I would resist is the framing that radiomics failed. It did not. It went through the normal arc of a quantitative biomarker field: an exciting early literature, a reproducibility reckoning, standardisation, and a slower, more credible second phase. That last phase is where it currently is, and it is the phase where useful things get built.
Tags
Taresh Sharan