A practical primer on radiomics and prognostic modelling

Radiomics turns medical images into measurable data. Instead of treating a CT, MRI or PET scan as a picture interpreted only through visual pattern recognition, it extracts quantitative descriptors of shape, intensity, texture and spatial relationships. These variables can then be studied alongside pathology, treatment details and patient outcomes.

The method sits at the intersection of imaging science, statistics, machine learning and clinical medicine. For medical physicists, radiologists, oncologists and data scientists, the important lesson is that a radiomics result is shaped long before a model is trained. Scanner settings, segmentation choices, feature definitions and validation design can all influence whether a promising association is clinically useful.

What radiomics measures

A radiomics workflow usually begins by identifying a region of interest, such as a tumour, organ or treatment volume. Software then calculates engineered imaging features. First-order statistics describe the distribution of voxel intensities, while shape features capture volume, surface area, compactness and irregularity. Texture measures estimate local variation and spatial organisation within the region.

Some projects also use wavelet filters, Laplacian-of-Gaussian transforms or other image decompositions to reveal patterns at different spatial scales. These derived features can represent heterogeneity that is difficult to judge consistently by eye. In oncology, the underlying hypothesis is that imaging heterogeneity may reflect cellular density, necrosis, vascularity, hypoxia or treatment resistance.

Radiomics is not automatically a form of artificial intelligence. A study may use a small set of predefined features in a Cox proportional hazards model, or thousands of variables in a machine-learning pipeline. The scientific value depends on reproducibility, biological plausibility, transparent reporting and a meaningful clinical endpoint.

Acquisition shapes the signal

Image acquisition determines what the algorithm can see. Field strength, tube voltage, reconstruction kernel, slice thickness, contrast timing, respiratory motion and scanner model may all alter feature values. A texture feature measured on a thin-slice CT reconstructed with a sharp kernel may not be directly comparable with the same feature from a smoother, thicker reconstruction.

A robust protocol therefore records acquisition metadata and seeks consistency wherever possible. In Australia, a project spanning a tertiary hospital in Melbourne, a private imaging provider in Sydney and a regional service in Queensland may encounter several scanner vendors and different local protocols. The public-private split in Australian healthcare makes this kind of heterogeneity common rather than exceptional.

Useful acquisition checks include:

• Confirm voxel spacing, slice thickness and reconstruction settings.
• Record scanner manufacturer, model, software version and contrast protocol.
• Assess motion, metal artefact, truncation and incomplete anatomical coverage.
• Test feature stability across repeat scans or controlled phantom studies.
• Document every preprocessing step before model development begins.

The everyday clinical workflow matters as well. A busy department may prioritise a fast scan, a familiar reconstruction and a practical reporting queue over research-perfect uniformity. Radiomics studies need to accommodate those realities rather than assuming that every patient has been imaged under identical conditions.

Segmentation and feature engineering

Segmentation is often the largest source of variation after acquisition. A contour may be drawn manually, produced by a semi-automatic tool or generated by a deep-learning model. Differences in tumour boundaries can change volume, margins and texture calculations, especially when lesions are small, irregular or close to organs with similar intensity.

Researchers should define the segmentation protocol in operational terms. They need to state whether necrotic areas, surrounding oedema, invasive extensions or treatment cavities are included. Interobserver testing, contour perturbation and sensitivity analysis can show whether a claimed imaging signature survives reasonable clinical disagreement.

Feature engineering should be conservative and traceable. Intensity normalisation, resampling and discretisation can improve comparability, but each operation introduces assumptions. A feature that appears predictive only after several undocumented transformations may be impossible to reproduce. The Image Biomarker Standardisation Initiative provides useful terminology and reference definitions for many common features.

Dimensionality reduction is often necessary because radiomics datasets contain many more candidate variables than patients. Correlation filtering, principal component analysis or penalised regression can reduce redundancy. Feature selection must occur inside the training process, however; selecting variables using the full dataset before cross-validation allows information from the test folds to leak into the model.

Building a prognostic model

A prognostic model estimates an outcome such as overall survival, recurrence, progression or toxicity. The target must be defined precisely, including the time origin, censoring rules and follow-up window. For time-to-event outcomes, suitable approaches include Cox regression, random survival forests and neural survival models. Classification methods may be appropriate for a clearly defined fixed-time endpoint, such as two-year recurrence.

Clinical variables should be considered alongside radiomic features. Age, performance status, stage, molecular markers, treatment regimen and comorbidities may explain outcome variation that imaging alone cannot capture. A model that combines complementary information should be compared with a clinical-only baseline and, where relevant, an established scoring system.

A sound modelling workflow separates development from evaluation. Training data are used to fit preprocessing and model parameters. Internal validation estimates optimism, while an independent external cohort tests transportability. Calibration is as important as discrimination: a model may rank patients correctly yet systematically overestimate or underestimate absolute risk.

Common modelling safeguards include:

• Predefine the outcome, prediction horizon and analysis population.
• Keep feature selection and imputation within resampling loops.
• Report discrimination, calibration and clinically relevant decision measures.
• Examine missing data, class imbalance and competing risks.
• Preserve a locked model before external testing.

The final product should be interpretable enough for clinical review. A radiomics score that cannot be explained, reproduced or connected to a treatment decision is unlikely to gain trust, regardless of its area under the curve.

Validation in Australian practice

External validation should reflect the setting where the model will be used. A model developed from one Melbourne cancer centre may perform differently in Perth or Adelaide because of population characteristics, referral patterns, imaging equipment and treatment pathways. Validation across Australian cities, regional hospitals and private providers can expose these differences early.

Transportability also depends on the local market. Australian health services may use different PACS, radiology information systems, cloud arrangements and data-export processes. Smaller regional services may have fewer research staff and less consistent access to advanced imaging. A model requiring a particular sequence or reconstruction may therefore be impractical outside a well-resourced metropolitan centre.

Professional development supports careful evaluation. Researchers presenting methods, results and limitations to a specialist audience can place their work within the wider medical physics community, while the archived CCPM examinations page illustrates the profession’s emphasis on structured competence and formal standards. Radiomics should be treated with the same discipline as any other clinical technology.

A useful validation programme may include repeatability testing, temporal validation, geographic validation and prospective silent deployment. In silent deployment, the algorithm generates predictions without influencing care. Its outputs can be checked for missing inputs, processing failures, calibration drift and unexpected subgroup effects before clinicians rely on them.

Governance and clinical translation

Radiomics uses sensitive health information, so governance must be designed into the project. Australian researchers need to consider the Privacy Act 1988, the Australian Privacy Principles, state or territory health-record requirements and local Human Research Ethics Committee processes. Data linkage, cross-border cloud storage and secondary use of archived scans require clear approvals and documented access controls.

Consent and de-identification deserve careful attention. Removing names does not always eliminate re-identification risk, particularly when imaging is linked with dates, rare diagnoses or genomic data. Data dictionaries, audit trails, role-based permissions and retention rules help maintain accountability. If a commercial vendor is involved, contracts should address model ownership, security, updates and responsibility for errors.

Clinical translation also requires a defined decision pathway. The team should specify who reviews the output, how uncertainty is displayed, what action follows a high-risk estimate and when the result should be ignored. The algorithm should support professional judgement rather than quietly replacing it. Australian implementation may also involve Therapeutic Goods Administration considerations if software functions as a regulated medical device.

The history of scientific meetings shows why governance and collaboration matter: discussions about future venues and professional priorities are captured in the archive’s account of conference location planning. Radiomics benefits from the same cross-disciplinary exchange among physicists, clinicians, statisticians, software specialists and patients.

A credible project leaves a reproducible record of data selection, preprocessing, segmentation, feature extraction, model fitting and validation. Reporting standards such as IBSI-aligned feature definitions and transparent radiomics checklists can make that record auditable. The strongest studies are often those that state clearly where the method fails, which patients were excluded and what evidence is still missing.

Begin with a clinically important question, a stable imaging protocol and an outcome that can be measured reliably. Build the pipeline so another Australian service could reproduce it, test the model against a clinical baseline, and seek independent validation before allowing a radiomics prediction to influence patient care.