Segmentation is the starting point. Radiomics is what comes next — extracting hundreds of quantitative features from your tumor segmentation masks that capture patterns invisible to the human eye, then using them to predict molecular markers, survival, and treatment response without a biopsy.
Traditional radiology is qualitative: a radiologist looks at an image and describes what they see. Radiomics is quantitative: algorithms extract hundreds to thousands of numerical features from medical images that capture tumor characteristics invisible to the human eye — subtle texture patterns, intensity distributions, and shape irregularities that correlate with molecular biology, treatment response, and patient outcomes.
The core premise is simple: images contain far more information than doctors can perceive visually. A glioblastoma that appears homogeneous to a radiologist may have texture features that reveal heterogeneous cellular architecture, or intensity patterns that correlate with IDH mutation status. Radiomics turns these hidden patterns into data.
Standardized MRI acquisition (T1, T1ce, T2, FLAIR, optionally DWI/perfusion). Acquisition parameters affect feature values — this is where reproducibility challenges begin.
Delineate tumor sub-regions (ET, TC, WT, peritumoral edema) using nnU-Net or manual annotation. This is where your BraTS skills directly feed the radiomics pipeline. Segmentation quality is the single biggest source of feature variability.
Resample to isotropic voxels (1mm³), normalize intensities (Z-score for MRI), apply intensity discretization (bin width typically 25–50 for MRI). These settings dramatically affect features and must be standardized.
PyRadiomics extracts 100–2,000+ features per region: shape, first-order histogram, texture (GLCM, GLRLM, GLSZM, NGTDM), and higher-order (wavelet, LoG). Each feature captures a different aspect of tumor biology.
With more features than patients, dimensionality reduction is critical to avoid overfitting. Methods: remove low-variance and highly correlated features, apply LASSO or recursive feature elimination, validate stability across resampled subsets.
Train classifiers (logistic regression, random forest, SVM, XGBoost) or survival models (Cox regression) on the selected features. Validate on external, independent datasets — not just cross-validation on the training set.
Capture 3D tumor geometry: volume, surface area, sphericity, elongation, compactness, maximum 3D diameter. These are independent of image intensity — they describe the tumor’s physical form. Biologically, irregular shapes and low sphericity often indicate aggressive, infiltrative growth. Shape features are the most sensitive to segmentation quality: a slight boundary shift changes volume and surface area.
Describe the distribution of voxel intensities without considering spatial relationships: mean, median, variance, skewness, kurtosis, entropy, energy, percentiles (10th, 90th). These capture overall tumor density and cellularity. High entropy suggests heterogeneous tissue; skewed distributions may indicate necrosis or hemorrhage. These features are more robust to segmentation variability than shape or texture features.
Quantify spatial relationships between neighboring voxels, capturing intratumoral heterogeneity:
GLCM (Gray Level Co-occurrence Matrix): contrast, correlation, energy, homogeneity — how often pairs of voxel intensities occur together. High GLCM contrast indicates abrupt intensity transitions (tumor-necrosis boundaries).
GLRLM (Gray Level Run Length Matrix): captures how long continuous runs of the same intensity extend — long runs suggest homogeneous regions, short runs indicate texture.
GLSZM (Gray Level Size Zone Matrix): measures the size of connected zones of similar intensity. NGTDM: captures coarseness and busyness of texture patterns.
Derived from filtered images that enhance specific patterns: wavelet decomposition (8 sub-bands capturing high/low frequency patterns in each dimension) and Laplacian of Gaussian (LoG) filters (with σ from 1–5mm, capturing features at different spatial scales). Apply first-order and texture features to each filtered image, multiplying the total feature count. The IBSI 2024 standardization extended to these convolutional filters for cross-platform reproducibility.
PyRadiomics is the open-source Python package for IBSI-compliant radiomic feature extraction. It accepts NIfTI images and segmentation masks and outputs a structured feature vector. Here’s how to extract features from your BraTS segmentations:
from radiomics import featureextractor
import pandas as pd
import nibabel as nib
import numpy as np
# Configure extractor
params = {
"binWidth": 25, # Intensity discretization
"resampledPixelSpacing": [1, 1, 1], # Isotropic 1mm
"normalize": True, # Z-score normalize
"imageType": {
"Original": {},
"LoG": {"sigma": [1.0, 3.0, 5.0]},
"Wavelet": {}
}
}
extractor = featureextractor.RadiomicsFeatureExtractor(params)
# Extract features from enhancing tumor on T1ce
image_path = "patient_001_T1ce.nii.gz"
mask_path = "patient_001_seg.nii.gz"
# Label 3 = enhancing tumor (or 4 in original BraTS)
result = extractor.execute(image_path, mask_path, label=3)
# Convert to DataFrame
features = {k: v for k, v in result.items()
if not k.startswith("diagnostics")}
df = pd.DataFrame([features])
print(f"Extracted {len(features)} features")
# Typically: ~100 original + ~800 wavelet + ~300 LoG = ~1,200 total
This is radiomics’ most impactful clinical application: predicting tumor molecular markers that normally require an invasive biopsy, using only MRI-derived features. For patients where biopsy is risky (deep-seated tumors, elderly patients), radiomics could provide critical molecular information non-invasively.
The most clinically important molecular marker for glioma classification and prognosis. Meta-analyses report pooled sensitivity 86.7% and specificity 82.8% for radiomics-based IDH prediction, with individual studies achieving AUC 0.80–0.99. The most predictive features include texture from T2-FLAIR (capturing the T2-FLAIR mismatch sign), first-order features from T1ce, and shape features. A multicenter radiomics nomogram achieved AUC 0.891 in training and 0.881 in external validation. Deep learning models reach sensitivity 0.80 and specificity 0.85, and hybrid approaches combining both achieve AUC up to 0.98.
Predicts response to temozolomide chemotherapy — one of the most actionable molecular markers. Radiomics achieves pooled AUC 0.77 (sensitivity 77.5%, specificity 71.4%). Multiregional models combining features from enhancing tumor, necrotic core, and edema perform best (AUC 0.84–0.88 in validation). The most informative features come from FLAIR texture and T1ce first-order statistics. A novel multimodal framework combining habitat radiomics, deep learning, and conventional radiomics pushed performance to AUC 0.98.
Defines oligodendroglial tumors and predicts chemotherapy sensitivity. Radiomics achieves sensitivity 0.75, specificity 0.82 with AUC 0.71–0.99. Thin-slice 3D MRI-based models achieve AUC 0.94 in training and 0.89 in validation. Combining the T2-FLAIR mismatch sign with radiomics improves AUC from 0.77 (mismatch alone) to 0.88. Knowledge-guided radiomics incorporating visual features (enhancement pattern, calcification) shows better cross-center generalizability than purely data-driven approaches.
Radiomic signatures predict overall survival with C-indices of 0.65–0.89 in glioblastoma. An 11-feature radiomic signature stratified patients into high and low-risk groups with hazard ratios of 4.33 for OS and 2.43 for PFS. Habitat radiomics — analyzing features from tumor sub-regions separately — achieves the best performance (C-index 0.89) because different compartments capture different biological processes. Combining radiomics with clinical data (age, performance status, extent of resection) and molecular markers (IDH, MGMT) yields C-indices of 0.73–0.77 in integrated models. RNA sequencing pairing has revealed that specific radiomic features correlate with immune regulation, proliferation, and treatment response pathways.
Pre-treatment radiomic features predict response to chemoradiation with AUC up to 0.92 for residual gliomas. Peritumoral radiomics combined with clinical features predicts 12-month progression-free survival with AUC 0.75. Integration of dosiomics (features from the radiation dose distribution) with pre-treatment MRI radiomics improves survival stratification to AUC 0.96 — a powerful combination of “what does the tumor look like?” with “how was the tumor treated?”
Instead of extracting features from a single scan, delta-radiomics computes changes in features between timepoints. This captures the dynamic treatment response that single-timepoint features miss. Delta-features improve treatment assessment AUC from 0.65–0.78 to 0.89 when using 1–2 week post-treatment changes. An 8-year retrospective study achieved AUC 0.95–0.99 for predicting glioma recurrence using combined pre- and post-operative delta-habitat radiomics. The optimal window is 1–2 weeks post-treatment, capturing early biological changes before morphological progression is visible. This directly complements the RANO longitudinal monitoring you learned in Week 11.
Week 11 introduced the pseudoprogression problem. Radiomics is one of the most promising tools for solving it. Texture and intensity features from conventional and advanced MRI capture microstructural differences between treatment effects and true tumor that are invisible on standard visual assessment.
Features from standard T1ce, T2, FLAIR, and DWI/ADC sequences achieve accuracy 73–90%, AUC 0.80–0.96 for distinguishing pseudoprogression from true progression. Multi-sequence models using 6 MRI sequences with radiomics achieved AUC 0.88 with high robustness. Machine learning combining radiomic features with MGMT status and clinical factors reaches AUC 0.80.
The strongest discriminators come from perfusion MRI. A landmark multicenter study found perfusion-derived radiomic features identified pseudoprogression with AUC 0.90, sensitivity 91%, specificity 88%. Incorporating diffusion and perfusion into radiomics models improves AUC from 0.76–0.80 (conventional only) to 0.90. Combined quantitative susceptibility mapping (QSM) + arterial spin labeling (ASL) + diffusion features achieved AUC 0.95. The biological basis: pseudoprogression has reduced blood volume (low rCBV) and increased diffusivity (high ADC) compared to true progression.
The traditional radiomics pipeline (handcrafted features + ML classifier) and deep learning (end-to-end CNNs) are converging. Deep radiomics uses features extracted from intermediate CNN layers as descriptors, combining the interpretability of handcrafted features with the representation power of deep learning.
A head-to-head comparison found deep learning achieved AUC 0.89 vs radiomics AUC 0.86 for molecular subtyping, with feature visualization showing weak correlation between the two feature types — meaning they capture complementary information. Hybrid approaches combining both consistently outperform either alone: AUC 0.946 for glioma grading (vs 0.891 radiomics-only, 0.903 deep-only), and AUC 0.98 for IDH prediction using deep radiomics nomograms. Different combinations excel for different tasks: habitat + DenseNet features for grading, habitat + radiomics for IDH prediction. The lesson: don’t choose between radiomics and deep learning — combine them.
One of the most exciting radiomics findings is that features extracted from the peritumoral edema zone — the region surrounding the visible tumor — contain information about tumor biology that intratumoral features alone cannot capture. This makes biological sense: peritumoral edema contains infiltrating tumor cells, reactive astrocytes, and disrupted vasculature that reflect the tumor’s invasive potential.
Adding peritumoral features to intratumoral features improves glioma grading, MGMT prediction (AUC 0.84–0.88 in multicenter validation), and progression-free survival prediction (AUC 0.75). A 2025 study demonstrated that radiomics-based quantification of tumor infiltration in the non-enhancing peritumoral region on post-operative MRI was independently associated with survival in glioblastoma. This has a practical implication for your BraTS segmentations: the “whole tumor” region (which includes peritumoral edema) is not just a less precise label — it captures biologically meaningful information about tumor invasiveness.
Radiomics has a reproducibility problem that threatens clinical translation. Features are sensitive to everything: scanner type, acquisition parameters, reconstruction algorithms, and segmentation quality. If you extract features from the same tumor scanned on two different MRI machines, you may get significantly different values.
Only 4 of 17 features achieved excellent reliability (ICC > 0.9) across all software platforms. When restricted to IBSI-compliant software, this improved to 15 of 17. IBSI compliance improved cross-platform feature agreement from <35% to >95%. Different segmentation methods can alter feature values by >30%. Most published studies lack external validation (<30% include external cohorts), prospective design, or standardized preprocessing.
IBSI Standardization: The Image Biomarker Standardization Initiative defined 169 features with benchmark values for cross-platform validation. Always use IBSI-compliant software (PyRadiomics is compliant). ComBat Harmonization: Statistical method for removing batch effects between scanners/institutions, adapted from genomics. Apply ComBat before building models on multi-center data. Standardized Preprocessing: Fix resampling (1mm isotropic), bin width (25–50), normalization (Z-score for MRI), and document everything. Feature Stability Analysis: Assess which features are reproducible on your data before building models. Discard unstable features even if they seem predictive.
The Radiomics Quality Score (RQS) evaluates 16 components of study quality. The IBSI ensures feature computation reproducibility. TRIPOD-AI guidelines standardize reporting of prediction models. The 2025 ESR practice recommendations mandate IBSI compliance, standardized preprocessing, and transparent reporting for publication. If your radiomics study doesn’t meet these standards, it won’t be trusted — and shouldn’t be.
This is where the entire 12-week curriculum comes together. Segmentation quality directly impacts every radiomic feature value and every downstream clinical prediction. Shape features are most sensitive (a slight boundary shift changes volume and surface area). First-order features are more robust but still affected by which voxels are included. Texture features in between.
The good news: AI segmentation produces radiomic features equivalent to manual expert segmentation for clinical endpoints. Automated nnU-Net segmentation with Dice >0.85 for tumor core and >0.75 for enhancing tumor produces radiomic models maintaining AUC >0.90 for molecular classification. One study found automated segmentation-derived radiomics achieved AUC 0.92 vs 0.89 for manual — the automated pipeline was actually slightly better, likely due to reduced inter-observer variability.
examples/ directory has complete workflows for feature extraction from NIfTI images with segmentation masks.