From image to measurement table

The opening described a number as the output of a system. We can now follow that system. A CT and a region of interest may produce a table containing a volume, an attenuation summary, or many texture and shape values, but the table is not a transparent transcription of the patient. It is the result of a sequence of decisions. A useful first diagnostic is to write the sequence in full: acquisition/reconstruction -> region/segmentation -> preprocessing -> geometry and resampling -> intensity discretisation -> feature extraction -> robustness analysis -> inference and model.

The terms used along this path should not be allowed to blur together. In this chapter, quantitative imaging is the umbrella for numerical measurements derived from images, such as volume, diameter, or an attenuation summary. Radiomics is used for the higher-dimensional practice of extracting many engineered image features and analysing them statistically. An imaging biomarker is a stronger claim: extraction alone is insufficient. A feature earns that label only when it has technical validation, a defined biological, process, or response interpretation, and clinical validation matched to its intended use. Calling a feature a biomarker because a software package produced it confuses extraction with validation. The terminology is useful precisely because it separates a number, a feature collection, and a validated interpretation.

Following the chain

The first link is acquisition and reconstruction. The scanner, protocol, contrast phase, dose, field of view, reconstruction method, kernel, and slice thickness help determine the image on which the later measurement operates. Two reconstructions from one acquisition are therefore two image inputs, even when they show the same patient and lesion. Their roles should be recorded rather than hidden in a file name. In a comparison study, these variables may be design strata, covariates, or axes of a robustness experiment; the choice depends on the estimand.

The next link is the region. A mask determines which samples enter the measurement, and its reference geometry must correspond to the image. A shifted or differently drawn mask is not a small implementation detail: it defines a different population of voxels. Manual, semi-automatic, and model-generated segmentations can therefore be different measurement inputs. Boundary uncertainty should be treated as part of the measurement problem rather than hidden behind a single preferred contour.

Preprocessing and geometry then alter the samples before a feature is calculated. Filtering, intensity transforms, clipping, normalization, resampling grid, and interpolator all need an explicit place in the protocol. Resampling an image and resampling a label are different operations, as Chapter 4 showed. If an image is resampled on a defined grid and an intensity clip is then applied, reversing that order creates another transformation. That order can be a legitimate decision in a particular pipeline; it is not a universal recipe to apply without stating the purpose and the consequences.

Texture calculations commonly introduce another definition step by mapping a continuous intensity range to discrete grey levels. A feature calculated after one binning rule is not automatically the same feature calculated after another. The target spacing, the intensity range, the binning rule, and the mask jointly define the samples from which the matrix is built. The detailed texture families and discretisation examples belong to the next chapter, but the measurement-chain consequence belongs here: discretisation is part of what was measured, not cosmetic formatting applied after the measurement.

Feature extraction is the point at which an implementation applies mathematical definitions to those prepared inputs. The output can be a single value or a wide table. A table becomes useful only when its rows, columns, units, and input contract are known. A feature name without those conditions is a label for a value whose meaning is still incomplete.

Finally, robustness analysis and modelling must respect the information boundary. If a study uses outcome information or cohort variation to select features, tune a preprocessing rule, or decide which features are retained, that decision must be made using only the training data inside each validation split. An outcome- independent characterization may have a different placement, but it still needs a declared protocol and a clear estimand. The general rule is not that one named step always belongs in one location; it is that no step may use information from the held-out evaluation data while pretending that the evaluation remained untouched.

Using the chain to diagnose a surprising value

Suppose two studies report a feature with the same name but different distributions. The name alone does not establish a biological disagreement. Compare the image inputs, reconstruction, region definition, geometry, preprocessing, discretisation, implementation, and estimator in that order. If a setting differs, the values may be different quantities under different operational definitions. They can still be scientifically useful, but their comparison has to be framed around the intended question rather than assumed to be a direct comparison of tumour properties.

This chain is also a reporting discipline. A reproducible result does not require that every study use identical choices, but it does require that the choices be visible enough for a reader to know what was measured and for a later study to repeat or deliberately vary it. The next section narrows the same idea to the inputs that most directly determine a feature: image, mask, geometry, and preprocessing.

Sources

• Zwanenburg A, Vallières M, Abdalah MA, et al. The Image Biomarker Standardisation Initiative. Radiology. 2020;295(2):328-338. doi:10.1148/radiol.2020191145, PMID 32154773.
• Quantitative Imaging Biomarkers Alliance (QIBA), RSNA — official QIBA information.
• Pfaehler E, et al. Phys Imaging Radiat Oncol. 2021;20:69-75. doi:10.1016/j.phro.2021.10.007, PMID 34816024.
• Haarburger C, et al. Scientific Reports. 2020;10:12688. doi:10.1038/s41598-020-69534-6.
• Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement. BMJ. 2024;385:e078378. doi:10.1136/bmj-2023-078378.
• Bernett J, et al. Guiding questions to avoid data leakage in biological machine learning applications. Nature Methods. 2024;21:1444-1453. doi:10.1038/s41592-024-02362-y.