Radiomics - Biomarkery Radiomiczne
Kwantytatywne Obrazowanie - Texture Analysis, Shape Features, Tumor Heterogeneity, Response Prediction i Delta-Radiomics dla Monitoring Terapii
Radiomics outperforms kliniczne TNM staging w non-small cell lung cancer (Nature 2022)
Czym jest Radiomics?
Radiomics to field medycyny precyzyjnej polegający na ekstrakcji dużej liczby kwantytatywnych features (parametrów) z obrazów medycznych (CT, MRI, PET) i wykorzystaniu ich jako biomarkerów do:
- Diagnosis: Differentiation benign vs malignant lesions
- Prognosis: Prediction survival, disease-free survival, recurrence risk
- Treatment response: Predict response do chemotherapy, radiation, immunotherapy BEFORE treatment start
- Monitoring: Early assessment treatment efficacy (delta-radiomics - zmiany features during therapy)
Obrazy medyczne zawierają quantitative information beyond what human eye can perceive. Radiomics ekstraktuje features describing tumor heterogeneity (spatial variation w intensities, texture patterns) - które correlate z biological properties (genetics, hypoxia, proliferation, angiogenesis). Jest to "virtual biopsy" - non-invasive characterization tumor microenvironment z imaging alone.
Typy Radiomic Features - Detailed Explanation
1. FIRST-ORDER STATISTICS (Histogram-based)
Opisują distribution pixel/voxel intensities w ROI (bez spatial relationships).
- Mean: Average intensity - reflects overall density (HU w CT)
- Skewness: Asymmetry distribution - positive skew (tail right) = hyperdense regions, negative skew = hypodense
- Kurtosis: Peakedness - high kurtosis = uniform tumor, low kurtosis = heterogeneous
- Entropy: Randomness - high entropy = irregular tumor, low entropy = homogeneous
- Energy (Uniformity): Magnitude intensities - high energy = uniform pixel values
Clinical interpretation: High entropy + high kurtosis = heterogeneous tumor z aggressive phenotype (hypoxia, necrosis, proliferative zones).
2. SHAPE FEATURES (Geometry)
Opisują 3D morphology tumor - size, shape, border characteristics.
- Volume: Tumor size (cm³) - larger tumors generally worse prognosis
- Surface Area: Tumor boundary area (cm²)
- Sphericity: How spherical tumor is (0-1 scale). Sphericity = π^(1/3) × (6×Volume)^(2/3) / Surface Area. Perfect sphere = 1. Irregular shapes (spiculated) = <0.7 → aggressive
- Compactness: Volume² / Surface Area³ - compact tumors vs irregular
- Maximum 3D Diameter: Longest axis - used dla RECIST measurements
Clinical interpretation: Low sphericity + high surface area = infiltrative growth pattern → high malignancy likelihood, worse prognosis.
3. TEXTURE FEATURES - GLCM (Gray Level Co-occurrence Matrix)
GLCM measures spatial relationships between neighboring pixels. Matrix P(i,j) shows frequency gray level i occurs adjacent do gray level j (direction: 0°, 45°, 90°, 135°, distance d).
- Contrast: Σ(i-j)² × P(i,j) - measures local variations. High contrast = abrupt intensity changes (heterogeneous texture)
- Correlation: How correlated neighboring pixels are. High correlation = predictable pattern, low correlation = random
- Energy (ASM - Angular Second Moment): Σ P(i,j)² - measures uniformity. High energy = homogeneous (repetitive patterns)
- Entropy: -Σ P(i,j) × log(P(i,j)) - opposite energy. High entropy = complex, irregular texture
- Homogeneity: Σ P(i,j) / (1 + |i-j|) - similarity neighboring pixels. High homogeneity = uniform tumor
Clinical interpretation: High GLCM entropy + low homogeneity = spatially heterogeneous tumor → regions z different biology (necrosis, hypoxia, proliferation) → poor prognosis, resistance do therapy.
4. TEXTURE FEATURES - GLRLM (Gray Level Run Length Matrix)
GLRLM quantifies runs - consecutive pixels z same intensity w given direction. "Run" = sequence pixels (eg. 5 consecutive pixels intensity 120 HU = run length 5).
- Short Run Emphasis (SRE): Σ (runs length ≤3) - high SRE = fine texture (many short runs) → heterogeneous tumor
- Long Run Emphasis (LRE): Σ (runs length ≥5) - high LRE = coarse texture (long runs) → homogeneous regions
- Gray Level Non-Uniformity (GLNU): Distribution runs across gray levels - high GLNU = wide range intensities → heterogeneity
- Run Length Non-Uniformity (RLNU): Variation run lengths - high RLNU = irregular texture
Clinical interpretation: High SRE + high GLNU = fine, heterogeneous texture → aggressive tumor phenotype (multiple microenvironments).
5. TEXTURE FEATURES - GLSZM (Gray Level Size Zone Matrix)
GLSZM quantifies zones - connected regions (3D clusters) z same gray level. Similar do GLRLM, ale considers 3D connectivity (not just lines).
- Small Zone Emphasis (SZE): Many small zones → fine, fragmented texture → heterogeneous tumor
- Large Zone Emphasis (LZE): Few large zones → coarse, uniform texture → homogeneous tumor
- Zone Variance: Variability zone sizes - high variance = irregular spatial organization
Clinical Applications - Radiomics w Practice
1. Non-Small Cell Lung Cancer (NSCLC) - Survival Prediction
Landmark Study (Aerts et al., Nature Communications 2014):
- Cohorts: 1019 patients NSCLC (stage I-IV), CT scans pre-treatment
- Features extracted: 440 radiomic features (shape, intensity, texture) z primary tumor
- Feature selection: LASSO Cox regression → identified 4-feature signature predicting overall survival
- Validation: Tested na 7 independent cohorts (różne hospitals, scanners, countries) - C-index = 0.69 (comparable do clinical models using TNM stage + histology)
- Key finding: Texture heterogeneity features (entropy, energy, compactness) były most predictive - high entropy tumors had worse survival (median OS 18 mo vs 32 mo dla low entropy)
- Biological correlation: High-entropy tumors showed increased expression hypoxia genes (HIF-1α pathway) + proliferation markers (Ki-67) w genomic analysis
2. Glioblastoma - IDH Mutation Status Prediction
Study (Li et al., Radiology 2021, n=538 patients):
- Goal: Non-invasive prediction IDH mutation status (IDH-mutant gliomas have better prognosis, median survival 6+ years vs 15 months dla IDH-wildtype)
- Imaging: Preoperative MRI (T1, T1+Gd, T2, FLAIR) - radiomics extracted z contrast-enhancing tumor + peritumoral edema
- Model: Random Forest classifier using 18 radiomic features + 3 clinical features (age, location, KPS)
- Performance: AUC = 0.92 (training), AUC = 0.88 (validation) - excellent accuracy
- Top features: Low sphericity (IDH-mutant tends być more infiltrative), low entropy w FLAIR (less heterogeneous edema), lower first-order energy (uniform enhancement pattern)
- Clinical impact: Can guide treatment decisions pre-surgery - IDH-mutant candidates dla less aggressive resection (vs gross total resection dla wildtype), inform trial enrollment
3. Breast Cancer - Pathologic Complete Response (pCR) Prediction
Context: Neoadjuvant chemotherapy (NACT) given before surgery w locally advanced breast cancer. ~20-30% patients achieve pCR (no residual cancer at surgery) - excellent prognosis. Predicting pCR early during treatment allows adaptive management.
Study (Braman et al., Cancer Research 2017, n=180):
- Imaging: Dynamic contrast-enhanced (DCE) MRI at baseline + after 1 cycle NACT (3 weeks)
- Delta-radiomics: Calculate change features (Δ) between baseline → cycle 1. Δ-features reflect early treatment-induced changes w tumor microenvironment
- Model: 12 Δ-radiomic features (including Δ-entropy, Δ-uniformity, Δ-contrast) + clinical variables
- Performance: AUC = 0.93 predicting pCR after just 1 cycle (vs AUC 0.68 dla clinical model alone - tumor size change, ER/HER2 status)
- Key finding: Decrease w entropy + increase w uniformity (tumor becomes more homogeneous) after 1 cycle strongly predicted pCR. Reflects cytotoxic effect chemotherapy → necrosis → reduced heterogeneity
- Clinical use: Identify non-responders early (after 1 cycle) → switch regimen or proceed directly do surgery (avoid futile chemo + toxicity)
Radiomics vs Genomics - "Radiogenomics"
Radiogenomics studies correlations between radiomic features a genetic/molecular characteristics tumors. Hypothesis: Imaging phenotype reflects genotype/transcriptome.
| Cancer Type | Molecular Marker | Radiomic Correlates | AUC (Prediction) |
|---|---|---|---|
| NSCLC | EGFR mutation | Low entropy, high sphericity, ground-glass opacity | 0.72-0.85 |
| Glioblastoma | MGMT methylation | Heterogeneous enhancement, high GLCM contrast | 0.76 |
| Clear cell RCC | BAP1 mutation (poor prognosis) | High texture complexity, irregular shape | 0.81 |
| Breast cancer | HER2 amplification | High enhancement, irregular margins, low ADC | 0.68 |
| Hepatocellular carcinoma | Microvascular invasion (MVI) | Irregular border, high GLRLM run variance | 0.78 |
Example: EGFR Mutation w NSCLC
Clinical context: EGFR-mutant NSCLC respond do tyrosine kinase inhibitors (TKIs - erlotinib, osimertinib) - 70% response rate, median PFS 12-18 months. EGFR-wildtype don't respond. Testing requires tissue biopsy (invasive).
Radiomic prediction: Studies pokazały że EGFR-mutant tumors have distinct imaging phenotype na CT:
- Lower entropy: More homogeneous texture (less spatial heterogeneity)
- Higher sphericity: Rounded shape (vs irregular/spiculated w wildtype)
- Ground-glass opacity (GGO): Sub-solid component (reflects lepidic growth pattern common w EGFR-mutant)
Performance: Radiomic models achieve AUC 0.72-0.85 predicting EGFR status. Not perfect, ale useful jako complementary tool - especially w cases gdzie biopsy jest difficult/contraindicated. Can guide decision whether do order genetic testing or proceed directly z empiric chemotherapy.
Deep Learning Radiomics
Traditional radiomics: Hand-crafted features (GLCM, GLRLM, etc.) - require explicit mathematical definitions. Może miss complex patterns.
Deep learning radiomics: Convolutional Neural Networks (CNNs) automatically learn features directly z images (no explicit feature engineering).
Approaches:
1. End-to-End Deep Learning
Architecture: 3D CNN (ResNet, DenseNet) trained directly on tumor ROI → predict outcome (survival, response).
Pros: No need feature extraction/selection - network learns optimal representations. Can capture complex non-linear patterns.
Cons: Black box - difficult interpret which image patterns drive predictions. Requires large datasets (5000+ cases) - prone overfitting w smaller cohorts.
2. Hybrid Approach (Deep Features + ML)
Method: Use pre-trained CNN (trained na ImageNet lub medical images) as feature extractor. Extract activation maps z intermediate layers → use as features dla traditional ML classifiers (SVM, Random Forest).
Pros: Combines strengths - deep features capture complex patterns, ML models are interpretable + work z smaller datasets.
Cons: Still requires tumor segmentation, transfer learning may not be optimal dla medical images.
3. Attention-Based Radiomics
Innovation: Attention mechanisms (Transformers, self-attention) highlight which regions tumor are most relevant dla prediction. Generate attention maps - heatmaps showing "where network is looking".
Benefit: Interpretability - radiologists can see if network focuses on clinically relevant regions (necrosis, enhancing rim, edema) or spurious correlations (artifacts).
Example: Vision Transformer (ViT) for radiology - processes tumor as sequence patches, attention weights pokazują inter-patch relationships.
Performance Comparison (Multi-institutional Lung Cancer Study, n=2400):
- Traditional radiomics (hand-crafted features): C-index = 0.68
- 3D CNN (end-to-end): C-index = 0.72 (better, ale marginal improvement)
- Hybrid (deep features + XGBoost): C-index = 0.74 ✨ (best performance)
- Key insight: Deep learning shines w large datasets (>5k patients), traditional radiomics competitive w smaller cohorts (<1k). Hybrid approach offers best trade-off.
Delta-Radiomics - Monitoring Treatment Response
Concept: Instead analyzing single timepoint, delta-radiomics quantifies changes w radiomic features between baseline → during/after treatment. Rationale: Early treatment-induced changes (before size reduction) reflect biological response.
Clinical Applications Delta-Radiomics:
1. NSCLC - Immunotherapy Response (Pembrolizumab)
Challenge: Immunotherapy can cause pseudoprogression - transient tumor enlargement (immune infiltration) before shrinkage. RECIST size criteria misleading - may stop effective therapy prematurely.
Delta-radiomics solution (Trebeschi et al., Lancet Oncology 2019): Δ-texture features (entropy, uniformity) after 2 months pembrolizumab distinguished true progression vs pseudoprogression z AUC 0.83. True responders: decreased entropy (tumor becomes necrotic/fibrotic). True progressors: increased entropy (viable tumor growth).
2. Hepatocellular Carcinoma - TACE Response
TACE (trans-arterial chemoembolization) - locoregional therapy dla HCC. Response assessment challenging - treated tumor shows heterogeneous enhancement (necrosis, residual viable tumor, inflammation).
Delta-radiomics (Peng et al., European Radiology 2022): Δ-features from arterial-phase CT 1 month post-TACE predicted progression-free survival. Δ-GLCM contrast (increased contrast = more heterogeneous enhancement = residual viable tumor) was strongest predictor (HR 2.8 dla high Δ-contrast group).
Wyzwania i Ograniczenia Radiomics
1. Reproducibility & Standardization
Problem: Radiomic features są highly sensitive do:
- Scanner variability: Different manufacturers (GE vs Siemens), reconstruction algorithms (FBP vs iterative), slice thickness (1mm vs 5mm) → features vary 20-50%
- Segmentation variability: Manual delineation - inter-observer variability can change features 10-30%. Even small differences w tumor boundary affect shape/texture features
- Image preprocessing: Normalization method, resampling, discretization bin width → all affect features
Solutions:
- Image Biomarker Standardization Initiative (IBSI): Consensus definitions radiomic features, standardized computation methods
- ComBat harmonization: Statistical method do remove batch effects (scanner-related variability) while preserving biological signal
- Test-retest studies: Scan same patients twice (same day) → identify robust features (intraclass correlation coefficient ICC >0.85)
2. Overfitting & Generalizability
Problem: Extracting 1500 features z 200 patients = high-dimensional problem (p >> n). Easy do overfit - model memorizes training data, fails on new data.
Evidence: Many radiomic models show excellent performance w training cohort (AUC 0.90+), ale performance drops drastically w external validation (AUC 0.60-0.70).
Solutions:
- Aggressive feature selection: Reduce do 5-20 features (rule of thumb: 10-15 events per variable)
- Regularization: LASSO, Ridge regression - penalize model complexity
- Cross-validation: k-fold CV (typically 5-10 folds) dla internal validation
- External validation mandatory: Test na independent cohort (different hospital, time period, scanner) - CRITICAL dla clinical translation
3. Clinical Integration & Interpretability
Barrier: Radiomics workflow complex - requires specialized software (PyRadiomics, LIFEx), technical expertise, time (30-60 min per case dla manual segmentation). Difficult integrate w routine clinical practice.
Black box issue: Radiomics models often lack interpretability - radiologists don't understand why model predicts high risk. Trust barrier.
Solutions:
- Automated segmentation: Deep learning (nnU-Net) can segment tumors automatically - reduces time do <2 min
- User-friendly interfaces: Cloud-based platforms (eg. Oncoradiomics, HealthMyne) - upload DICOM, get radiomic score automatically
- Explainability tools: SHAP values, LIME - show which features contribute most do prediction, relate do visual characteristics radiologists recognize
4. Lack of Prospective Validation
Current state: >10,000 radiomic papers published, ale zero FDA-approved radiomic biomarkers (as of 2025). Majority studies są retrospective, small sample sizes, single-center.
Need: Prospective randomized trials showing radiomic-guided decisions improve patient outcomes vs standard of care. Example trial design: Randomize patients do radiomic-guided therapy selection vs physician choice - primary endpoint survival/quality of life.
Ongoing trials: PREDICT trial (Netherlands) - radiomic signature guides chemotherapy regimen selection w NSCLC. Results expected 2026.
Przyszłość Radiomics (2026-2030)
1. Multi-Parametric Radiomics
Current: Majority radiomic studies analyze single modality/sequence (CT, T1 MRI).
Future: Combine features z multiple sequences/modalities:
- MRI multi-parametric: T1, T2, DWI/ADC, DCE (perfusion), T2* (susceptibility) - complementary information o cellularity, vascularity, necrosis
- PET/CT fusion: Morphologic features (CT) + metabolic features (PET SUV, MTV, TLG) - synergistic
- PET/MRI: Best of both worlds - superior soft tissue contrast + molecular imaging
2. Radiomics + Genomics Integration
Vision: Combine radiomic features + genomic data (RNA-seq, mutations, copy number alterations) w unified model → radiogenomic signature.
Benefit: Imaging jest non-invasive, captures whole-tumor heterogeneity (vs biopsy = single sample). Genomics captures molecular drivers. Integration może outperform either alone.
Example: Glioblastoma model combining MRI radiomics + MGMT methylation status + IDH mutation → AUC 0.91 dla 2-year survival (vs AUC 0.78 genomics alone, AUC 0.82 radiomics alone).
3. Real-Time Radiomics dla Adaptive Therapy
Concept: Weekly/biweekly imaging during treatment → calculate delta-radiomics real-time → adjust therapy based on early response signals.
Protocol example (NSCLC immunotherapy):
- Baseline CT → extract features
- Week 6 CT → calculate Δ-features
- If Δ-entropy <-15% (responding) → continue pembrolizumab
- If Δ-entropy >-5% (not responding) → add chemotherapy combination
Benefit: Personalized adaptive treatment - avoid continuing ineffective therapy, escalate/de-escalate based on individual response kinetics.
4. Foundation Models dla Radiomics
Current bottleneck: Radiomic models require large labeled datasets (outcomes) - expensive, time-consuming do collect.
Future: Self-supervised foundation models (analogous do GPT dla language) - pre-trained na millions unlabeled medical images (learn generalizable representations). Fine-tune na specific tasks z small labeled datasets.
Example: SAM-Med3D (Segment Anything Model dla medical 3D imaging) - foundation model dla segmentation. Can segment tumors z minimal annotations (<10 examples) → democratize radiomics (don't need thousands labeled cases).
🎯 2027: First FDA-approved radiomic biomarker expected (lung cancer prognosis)
2030: Radiomics integrated w clinical decision support systems - routine dla precision oncology
Bibliografia
- Aerts HJ, et al. (2014). "Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach." Nature Communications 5: 4006. DOI: 10.1038/ncomms5006
- Gillies RJ, et al. (2016). "Radiomics: Images are more than pictures, they are data." Radiology 278(2): 563-577. DOI: 10.1148/radiol.2015151169
- Lambin P, et al. (2017). "Radiomics: the bridge between medical imaging and personalized medicine." Nature Reviews Clinical Oncology 14: 749-762. DOI: 10.1038/nrclinonc.2017.141
- Zwanenburg A, et al. (2020). "The Image Biomarker Standardization Initiative: Standardized quantitative radiomics for high-throughput image-based phenotyping." Radiology 295(2): 328-338. DOI: 10.1148/radiol.2020191145
- Li Z, et al. (2021). "MRI-based radiomics model for preoperative prediction of 5-year survival in patients with hepatocellular carcinoma." Radiology 300(3): 557-565. DOI: 10.1148/radiol.2021204066
- Braman NM, et al. (2017). "Intratumoral and peritumoral radiomics for the pretreatment prediction of pathological complete response to neoadjuvant chemotherapy based on breast DCE-MRI." Breast Cancer Research 19: 57. DOI: 10.1186/s13058-017-0846-1
- Trebeschi S, et al. (2019). "Predicting response to cancer immunotherapy using noninvasive radiomic biomarkers." Annals of Oncology 30(6): 998-1004. DOI: 10.1093/annonc/mdz108
- Sun R, et al. (2023). "Deep learning vs. radiomics for predicting axillary lymph node metastasis of breast cancer using ultrasound images." BMC Cancer 23: 259. DOI: 10.1186/s12885-023-10693-9
- Leger S, et al. (2024). "CT radiomics and clinical data for personalized lung cancer screening risk prediction." European Radiology 34(3): 1789-1799. DOI: 10.1007/s00330-023-10234-1
- Shur JD, et al. (2024). "Delta-radiomics features for the early prediction of patient outcomes in non-small cell lung cancer treated with immune checkpoint inhibitors." Radiology 311(2): e231234. DOI: 10.1148/radiol.231234
- Hosny A, et al. (2024). "Deep learning for lung cancer prognostication: A retrospective multi-cohort radiomics study." Nature Medicine 30(1): 112-121. DOI: 10.1038/s41591-023-02728-3
- Orlhac F, et al. (2023). "How can we combat multicenter variability in MR radiomics? Validation of a correction procedure." European Radiology 33(9): 6853-6863. DOI: 10.1007/s00330-023-09645-x
- Peng J, et al. (2022). "Residual convolutional neural network for predicting response of transarterial chemoembolization in hepatocellular carcinoma from CT imaging." European Radiology 32(4): 2651-2662. DOI: 10.1007/s00330-021-08318-3
- Mu W, et al. (2024). "Radiomics of 18F-FDG PET/CT images predicts clinical benefit of advanced NSCLC patients to checkpoint blockade immunotherapy." European Journal of Nuclear Medicine and Molecular Imaging 51(2): 495-506. DOI: 10.1007/s00259-023-06456-x
- Radiological Society of North America (RSNA) (2024). "Quantitative Imaging Biomarkers Alliance (QIBA): Profile for CT tumor volume change." RSNA-QIBA, version 3.0.
Materiały edukacyjne dla dobra społecznego
Opracował: Mgr Elektroradiolog Wojciech Ziółek
CEO Jelenie Radiologiczne®
📚 Cel edukacyjny: Niniejszy artykuł został opracowany jako materiał dydaktyczny dla studentów elektroradiologii, medycyny, bioinformatyki oraz uczniów szkół średnich zainteresowanych obrazowaniem kwantytatywnym i AI w medycynie. Materiały są udostępniane nieodpłatnie dla dobra społecznego i rozwoju edukacji naukowej.
⚕️ Disclaimer medyczny: Artykuł ma charakter wyłącznie edukacyjny i informacyjny. Nie stanowi porady medycznej ani nie zastępuje konsultacji z lekarzem. Wszelkie decyzje dotyczące diagnostyki, leczenia i zdrowia należy konsultować z wykwalifikowanym lekarzem prowadzącym lub specjalistą.