Radiomics
Quantitative Radiomics · Also known as: texture analysis, radiomics analysis, quantitative imaging biomarkers
Radiomics is a computational methodology that extracts large numbers of quantitative features from medical images (CT, MRI, PET) using automated image analysis and machine learning to discover imaging biomarkers associated with disease phenotype, prognosis, and treatment response. Developed by Lambin, Gillies, and colleagues in 2012, radiomics aims to decode the biology underlying visible imaging patterns, enabling personalized medicine through image-based phenotyping. It has emerged as a powerful tool in oncology for tumor characterization, prognosis prediction, and therapy response assessment.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Radiomics is indicated when characterization of spatial imaging patterns is clinically relevant to prognosis or treatment decision-making. Most developed in oncology (tumor grading, survival prediction, therapy response), radiomics is expanding into non-oncology (lung fibrosis, cardiac dysfunction, brain pathology). Radiomics requires high-quality segmentations, sufficient sample size for training/validation, and rigorous methodology to avoid overfitting. Radiomics is less applicable to simple diagnostic questions (presence/absence of disease) where visual interpretation suffices; it is most valuable in prognostic or phenotyping applications.
Strengths & limitations
- Quantitative objectivity: automated feature extraction eliminates inter-observer variability inherent in visual assessment, enabling standardized, reproducible biomarkers
- High information density: hundreds of features are extracted from each image; machine learning identifies feature combinations invisible to human interpretation
- Prognostic accuracy: radiomics models often outperform clinical scoring systems in predicting outcome (survival, recurrence, therapy response) in validation cohorts
- Personalized medicine: radiomics enables image-based patient stratification, guiding treatment selection and risk-adapted therapy
- Biologic insight: associations between imaging features and genomics/pathology unveil the biologic basis of imaging patterns
- Overfitting and non-reproducibility: radiomic models trained on small samples or with excessive feature tuning fail to generalize to independent cohorts; many published radiomics papers fail external validation
- Segmentation sensitivity: radiomic features are sensitive to segmentation accuracy; manual/semi-automatic segmentation introduces variability and bias
- Image protocol dependence: radiomics features depend on acquisition parameters (slice thickness, reconstruction kernel, MRI sequence, field strength); lack of standardization limits cross-site applicability
- Publication bias and multiple comparisons: emphasis on positive predictive associations without adequate multiple-comparison correction and pre-registration leads to false discoveries
- Lack of clinical adoption: radiomics models are rarely integrated into clinical workflows; regulatory approval and clinical utility remain limited
Frequently asked
What are the main types of radiomics features, and what do they represent?
Shape features (volume, sphericity) describe geometric properties. First-order statistics (mean, standard deviation, skewness, kurtosis) characterize intensity distributions. Textural features (GLCM, GLRLM, LBP) describe spatial relationships: GLCM represents pixel-pair relationships at various distances (coarseness); GLRLM describes runs of same-intensity pixels (homogeneity); LBP characterizes local intensity patterns. Higher-order features (wavelets, Laplacian) capture multiscale information. Together, these features summarize tumor heterogeneity and complexity.
How do I avoid overfitting in radiomics model development?
Key safeguards: (1) Proper train-test split: use 70/30 or 80/20 split with randomization; (2) Cross-validation: k-fold (k=5-10) assesses generalization within the training set; (3) Feature selection: reduce feature count before model training (LASSO, univariate filtering); (4) Adequate sample size: aim for n > 100 and prefer n >= 200 when possible; (5) Regularization: use L1/L2 penalties (LASSO, Ridge regression) to shrink non-informative features; (6) External validation: test in an independent cohort from a different institution. Models that work on training data but fail external validation are overfit.
Does radiomics feature choice depend on tumor type or imaging modality?
Yes. Different tumor types and imaging modalities may benefit from different feature sets. Textural features are most useful in heterogeneous tumors (glioblastoma, lung cancer); shape features dominate in well-defined lesions (renal cell carcinoma). CT features emphasize density; MRI features leverage multiple sequences and relaxation times. PET features reflect metabolic activity. Optimal features are identified empirically in cohort-specific studies; generalization across tumor types/modalities is limited.
How do image acquisition parameters affect radiomics features?
Radiomics features are sensitive to slice thickness (affects shape features), reconstruction kernel (affects texture features), field strength (MRI), and sequence parameters. Thicker slices artificially reduce feature variance; different kernels (soft, standard, edge-enhancing) drastically change texture. Cross-scanner and cross-protocol radiomics models suffer degraded performance. Best practice: standardize acquisition protocols, use retrospective image harmonization (e.g., ComBat), or retrain models for each protocol.
What validation evidence should a radiomics model have before clinical use?
Minimal evidence: (1) Internal validation (cross-validation in the training cohort); (2) At least one independent external validation cohort from a different institution; (3) Comparison to clinical baseline (staging, conventional scores); (4) Adequate sample size (n > 100 per cohort); (5) Pre-registration of the analysis protocol (INPLASY, OSF) to prevent selective reporting. Ideal: (6) Prospective validation; (7) Multiple external validation cohorts; (8) Integration with genomic and clinical data; (9) Health economic analysis. Most published radiomics models lack proper external validation; caution is warranted.
Sources
- Lambin, P., Rios-Velazquez, E., Leijenaar, R., et al. (2012). Radiomics: extracting more information from medical images using advanced feature analysis. Nature Reviews Clinical Oncology, 9(12), 676-684. DOI: 10.1016/j.ejca.2011.11.036 ↗
- Gillies, R. J., Kinahan, P. E., Hricak, H. (2016). Radiomics: images are data. Radiology, 278(2), 563-577. link ↗
- Kumar, V., Gu, Y., Basu, S., et al. (2012). Radiomics: the process and the challenges. Magnetic Resonance Imaging, 30(9), 1234-1248. DOI: 10.1016/j.mri.2012.06.010 ↗
How to cite this page
ScholarGate. (2026, June 3). Quantitative Radiomics. ScholarGate. https://scholargate.app/en/medical-imaging/radiomics
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- CT Iterative ReconstructionMedical Imaging↔ compare
- DEXAMedical Imaging↔ compare
- OCT AngiographyMedical Imaging↔ compare
- PET Kinetic ModelingMedical Imaging↔ compare
- Quantitative Susceptibility MappingMedical Imaging↔ compare