Skip to content
Imaging Evidence Path

Harmonisation and the model handoff

The RIDER case characterised measurement behavior before an intervention was chosen. That order matters when a feature table is assembled from multiple scanners, sites, or acquisition protocols. A feature can differ between groups because the image-generating process differs, but a group can also differ in patient population, care pathway, disease spectrum, or outcome distribution. The observed batch difference may contain several of these components at once.

Harmonisation is an attempt to make measurements more comparable across a declared source of variation. It is not a generic cleaning step and it does not turn a heterogeneous dataset into a single neutral population. The useful question is more precise: what variation is being adjusted, what information may be removed with it, and which parts of the adjustment are allowed to influence a later model?

Acquisition differences enter the feature table

The same nominal lesion can produce different feature distributions when scanner, protocol, reconstruction, resolution, or other acquisition conditions change. The difference may appear in first-order values, shape through the segmentation and geometry pathway, or texture through spatial intensity relationships. A model trained on pooled data can use such differences as shortcuts if they are associated with the outcome in the development cohort.

The word batch names a grouping used in an analysis; it does not identify the cause of the grouping difference. A site may be associated with a scanner protocol and with a different patient mix. The site signal may then include technical nuisance, real population differences, care-pathway differences, or several of these at once. It is not safe to define every site-associated signal as non-biological before the estimand has been considered.

This is the setting in which harmonisation becomes relevant. The intervention must be tied to a stated target: reducing a known acquisition-related difference, making a particular comparison possible, or supporting a later predictive model. Those are different purposes. A descriptive analysis of measurement behavior should not be silently replaced by a harmonised table whose differences have already been altered.

ComBat adjusts distributions under assumptions

ComBat is a feature-level batch-effect adjustment originally developed for microarray data and later used in multi-centre imaging analyses. In broad terms, it estimates a batch-related transformation of feature distributions. The transformation can be useful when its assumptions match the study design, but it is not a button that removes every site effect.

A central assumption is that the batch effect can be separated from the biological or outcome-related structure that should be retained. If the batch is entangled with biology or outcome, adjustment can remove part of the signal of interest, leave residual bias, or make the resulting feature less representative of the intended estimand. A site with a different scanner and a different case mix is not an example in which the technical and biological components can be assumed independent merely because the site label is available.

ComBat acts on already extracted feature values rather than on acquisition physics or raw images. If a reconstruction has removed spatial content from an image, transforming the extracted values does not reconstruct the image or restore spatial content absent from those inputs. If a feature is sensitive to a reconstruction choice, harmonising its distribution may make the values look more comparable while hiding the raw sensitivity that the measurement study was meant to characterise. A harmonised feature is therefore a feature after an intervention, not the untouched baseline measurement.

Feature-level ComBat and image-level harmonisation or domain-adaptation methods are different method classes. Feature-level ComBat acts on extracted values; image-level methods intervene before extraction and carry method-specific assumptions about the image domain and what should be preserved. The original method is described by Johnson, Li, and Rabinovic, and the imaging-specific considerations are discussed by Orlhac et al.. The relevant design question is not whether ComBat is generally good or bad. It is whether the proposed adjustment is appropriate for the batches, biological structure, outcome, and intended use in this study.

Characterisation and model development are different phases

The distinction between descriptive characterisation and model development determines how a harmonisation step should be handled.

An outcome-independent descriptive analysis may ask how feature distributions vary across acquisition groups or how a defined feature behaves under a declared perturbation. Its purpose is to describe the measurement and diagnose a possible source of variation. It should be recorded as such, without presenting the result as evidence that a later predictive intervention will help.

An outcome-informed analysis asks a different question. If the outcome is used to choose a harmonisation rule, tune its parameters, retain particular features, or compare model variants, those choices belong to model development. Stability-based selection has the same status when its result determines which columns enter the model. The fact that a selection rule was motivated by measurement performance does not make it independent of model development once it selects model inputs.

Reporting guidance such as the CLEAR radiomics reporting checklist and TRIPOD+AI makes these development and evaluation steps part of the declared study design. These resources guide reporting; they do not provide empirical validation.

The validation boundary is the governing invariant. Any fitted or data-dependent transformation used to construct predictive inputs, including ComBat, scaling, feature selection, and outcome-informed stability selection, must be fitted using only the training information permitted by the design and then applied to the held-out data. This training/evaluation separation is explicit in TRIPOD+AI and in CLEAR’s reporting items for training-only reliability, scaling, reduction, model selection, and testing. These guidelines make the procedure inspectable; they do not validate the transformation or the model. If a transformation is fitted using the whole dataset before cross-validation, the held-out feature distribution has influenced the model inputs. The resulting evaluation no longer represents the declared training-only procedure.

This does not imply one universal order for every radiomics study. The appropriate placement depends on the estimand, whether the operation is descriptive or predictive, whether it acts on images or features, what reference data are available, and which sources of variation are being studied. A harmonisation step can be a separate measurement intervention, a component of model development, or part of another explicitly defined design. The role and information boundary must be fixed before the result is used to support a claim.

The RIDER-style baseline and a later harmonised analysis should consequently be treated as related but distinct measurements. First characterise the unadjusted behavior relevant to the question. Then, if harmonisation is justified, define a new intervention and re-measure under its declared contract. This preserves the ability to say what the intervention changed instead of confusing adjustment with the original measurement performance.

The model handoff is another claim boundary

After measurement performance and any separate intervention have been characterised, the feature table can be handed to a predictive analysis. That handoff does not mean that the columns have become independent, biological, or clinically useful.

Radiomics can produce candidate sets in which the number of features is much larger than the number of patients, expressed as p >> n. Shared inputs or related transformations can induce substantial dependence or correlation among candidates; they do not guarantee an observed correlation for every pair. The CLEAR radiomics reporting checklist makes feature reduction and model selection reportable parts of the pipeline, but it does not establish empirical validity. Model development therefore needs honest feature selection, an appropriate form of regularisation, and a validation design that includes every data-dependent choice. Harmonisation does not remove this dimensionality or selection problem.

The relevant comparison is also not simply whether an imaging model beats chance. Where the estimand calls for incremental value, compare it with a clinical tabular baseline containing variables such as age, stage, or histology. The CLEAR radiomics reporting checklist includes non-radiomic and combined comparisons as reporting considerations; that requirement does not itself demonstrate incremental value. The question is whether the imaging measurements add information beyond the clinical record under the same patient-level design.

An attractive internal predictive metric remains an internal result. It does not by itself establish that the feature measures the intended biology, that the result transports to another acquisition setting, or that using the model improves a clinical decision. Calibration, external evaluation, and clinical utility answer different questions from internal discrimination. TRIPOD+AI distinguishes these prediction-reporting dimensions; the guideline makes the claims inspectable but does not supply empirical validation. The model handoff carries those distinctions forward rather than replacing them with one score.

Three complementary evidence roles

Several standards and reporting resources support different layers of this work. They should be used for their respective questions rather than treated as interchangeable approval labels.

QIBA addresses measurement performance under a specified imaging protocol. Its role is to make questions about bias, precision, repeatability, and reproducibility concrete for a defined measurement use. The official QIBA information is the appropriate entry point for that layer.

IBSI addresses how radiomics features and filters are defined and computed, including reference calculations against which implementations can be checked. It helps determine what number an implementation is supposed to produce, but agreement with a computational definition does not by itself establish biological validity or clinical utility. The IBSI reference paper and the later IBSI 2 publication address these standardisation questions.

CLEAR addresses reporting of radiomics studies. Reporting guidance helps a reader reconstruct the cohort, measurements, processing, analysis, and claims well enough to assess what the study actually established. It cannot make an unspecified measurement reproducible or turn a weak validation design into clinical evidence. The CLEAR radiomics reporting checklist is described by Kocak et al..

These roles are complementary. QIBA can frame whether a measurement performs under declared conditions, IBSI can clarify how the feature or filter is computed, and CLEAR can make the study’s evidence inspectable. None of the three removes the need to state the estimand, preserve the validation boundary, or limit the final claim to what the design supports.

The next section returns to the ordering problem in a bounded case: Case study: measure before intervening examines why a baseline measurement should be preserved before harmonisation, filtering, or modelling is introduced, and how each intervention can be given its own question and comparison.

Sources

Adrian Tkacz, MD, MSc Computer ScienceAboutLinkedInORCIDXEmail

© 2026 Adrian Tkacz · BioMedical-IT
Original text and figures are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0), unless otherwise noted. Please credit Adrian Tkacz / BioMedical-IT and link to the original source. Third-party materials retain their respective rights and licenses.