Radiomics and Medical Imaging AI
Educational note: this page reflects public evidence last checked on 2026-06-04. Radiomics features are research and engineering biomarkers unless validated for a specific clinical task, population, scanner workflow, and decision point.
TL;DR
Radiomics extracts quantitative features from CT, MRI, PET, ultrasound, or pathology images: shape, intensity, texture, spatial heterogeneity, and filtered-image patterns. Those features can be studied as imaging biomarkers for prognosis, molecular association, segmentation QA, treatment response, and trial stratification. The scientific risk is overclaiming: most radiomics models are not automatically portable across scanners, sites, segmentations, reconstruction protocols, or patient populations. Treat radiomics as a measurement and validation problem, not as a shortcut around pathology, genomics, or clinical trials. Sources: [1], [2], [3]
1. The core concept
Medical images are not just pictures; they are structured arrays with acquisition metadata, reconstruction choices, anatomy, and measurement artifacts. A radiomics workflow converts a segmented region of interest into mathematical features that may correlate with tissue architecture, necrosis, cellularity, perfusion, hypoxia, or genotype.
The phrase may correlate is doing important work. A radiomic texture feature is not direct evidence of a mutation, immune state, or treatment sensitivity unless the model has been validated against an appropriate reference standard.
2. Typical workflow
- Acquisition and metadata capture - scanner, sequence/protocol, contrast, reconstruction, voxel spacing, and timing.
- Preprocessing - resampling, intensity normalization, discretization, registration, artifact control.
- Segmentation - tumor, organ, lesion, or subregion mask; manual, semi-automated, or deep-learning-assisted.
- Feature extraction - shape, first-order, texture, and filtered-image features, ideally using IBSI-compatible definitions.
- Modeling - leakage-safe train/validation/test split, feature selection inside cross-validation, calibration, uncertainty, and external validation.
- Clinical interpretation - decision curve, failure analysis, subgroup performance, and comparison with simpler baselines.
IBSI standardization exists because small implementation differences can change feature values. Reproducibility is part of the science, not a cosmetic software detail. Sources: [1], [2]
3. Feature families
| Feature family | What it measures | Main caveat |
|---|---|---|
| Shape | Volume, diameter, compactness, sphericity, surface area | Depends heavily on segmentation |
| First-order | Histogram statistics such as mean, variance, skewness, entropy | Sensitive to intensity scaling and acquisition |
| Texture | GLCM, GLRLM, GLSZM, NGTDM, GLDM patterns | Sensitive to discretization, voxel size, and noise |
| Filtered-image | Wavelet, Laplacian-of-Gaussian, local binary patterns | Easy to overfit when feature count is high |
| Deep imaging features | CNN or foundation-model embeddings | Less interpretable; needs large, representative data |
4. Evidence maturity
translational
Radiomics is mature as a research workflow and useful as an engineering surface for imaging QA, feature extraction, and hypothesis generation. It is not mature as a general-purpose clinical decision engine. A radiomics model should be treated as clinically meaningful only when the exact task has external validation, prospective or locked-model evaluation when appropriate, calibration, and comparison against current clinical standards.
5. Applications in oncology
- Radiogenomics - predicting molecular markers such as EGFR or IDH status from imaging is a research question; it does not replace tissue or validated molecular testing.
- Treatment-response modeling - baseline or longitudinal imaging may help stratify risk, but endpoint definitions and therapy context matter.
- Progression vs pseudoprogression - imaging models can support research and triage, but ambiguous cases still require multidisciplinary review.
- Segmentation and measurement QA - radiomics can expose unstable masks, scanner drift, and preprocessing differences.
- Clinical trial enrichment - imaging features may help define cohorts only when the assay is prespecified and reproducible.
6. Failure modes
- Training on one institution and assuming portability to another.
- Selecting features before cross-validation, causing leakage.
- Reporting high AUC without calibration, confidence intervals, or external validation.
- Treating a retrospective association as a diagnostic or companion test.
- Ignoring scanner, reconstruction, contrast timing, slice thickness, or segmentation variability.
- Calling a model "AI biomarker" without a locked assay and source-controlled preprocessing.
7. What technologists can build
- DICOM-to-feature pipelines with metadata capture and IBSI-compatible extraction.
- Segmentation QA dashboards that quantify inter-reader and model variability.
- Leakage-safe benchmark harnesses for radiomics models.
- Model cards for imaging biomarkers: scanner scope, preprocessing, endpoint, validation cohorts, failure modes.
- Data loaders that preserve acquisition metadata instead of flattening images into anonymous arrays.
See also
- AI and machine learning
- Biomarkers and companion diagnostics
- ML pitfalls in oncology
- Oncology data standards
References
- Zwanenburg A, Vallieres M, Abdalah MA, et al. The Image Biomarker Standardisation Initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology 2020;295:328-338. PMID 32154773. https://doi.org/10.1148/radiol.2020191145
- van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res 2017;77:e104-e107. PMID 29092951. https://doi.org/10.1158/0008-5472.CAN-17-0339
- Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-762. PMID 28975929. https://doi.org/10.1038/nrclinonc.2017.141