☢️ Radiomics Feature Extraction Tumor Phenotyping

Image Acquisition and the Foundation of Reproducible Radiomics

Radiomics is the high-throughput extraction of large numbers of quantitative features from standard-of-care medical images — CT, MRI, PET — that convert routinely acquired pixels into mineable, statistically analyzable data. Because texture and intensity features are exquisitely sensitive to acquisition parameters, the entire radiomics pipeline begins with a rigorous, standardized imaging protocol.

• 500–1800: Typical radiomics features/tumor (first-order + shape + texture)
• ≤1.5 mm: CT slice thickness (protocol) (IBSI-recommended for texture stability)
• up to 40%: Reconstruction kernel effect (feature value variation, same tumor)
• >15,000: Radiomics publications (2012–2024) (PubMed-indexed, exponential growth)

From pixels to mineable data — the radiomics hypothesis

The central hypothesis of radiomics, formalized by Lambin et al. (2012) and Aerts et al. (Nat. Commun. 2014), is that tumor heterogeneity — genetic, cellular, and microenvironmental variation within and between lesions — leaves a macroscopic imprint on the spatial pattern of voxel intensities visible on standard CT, MRI, or PET images. Because tumors are routinely imaged as part of standard clinical care, this signal is available essentially "for free," without additional biopsy or cost.

The radiomics workflow proceeds through five canonical steps: (1) image acquisition, (2) segmentation of the region/volume of interest, (3) feature extraction, (4) feature selection and model building, and (5) validation — each of which is a distinct, quantifiable source of variability that must be controlled for the resulting biomarker to be clinically trustworthy.

Unlike a radiologist's qualitative impression ("irregular, spiculated margin", "heterogeneous enhancement"), radiomics converts these visual impressions into hundreds of precise numerical descriptors — many of which capture texture patterns imperceptible to the human eye — that can be statistically correlated with underlying biology, treatment response, and prognosis.

Aerts et al. (2014, Nature Communications) analyzed 440 CT-derived radiomic features in 1,019 lung and head-and-neck cancer patients, finding a "radiomic signature" of 4 features that was prognostic for survival across independent cohorts and correlated with underlying gene-expression patterns — one of the founding studies of the field.

Acquisition parameters — the reproducibility bottleneck

Texture features are highly sensitive to technical acquisition parameters, which is why radiomics is often criticized for poor reproducibility across institutions unless carefully harmonized:

• Slice thickness: thinner slices (≤1.5 mm) preserve fine texture detail; features computed on 5 mm slices can differ substantially from the same tumor at 1 mm • Reconstruction kernel: sharp (bone/lung) kernels amplify high-frequency texture; smooth (soft-tissue) kernels suppress it — the same tumor can show markedly different GLCM entropy depending on kernel choice alone • Reconstruction algorithm: FBP vs. iterative vs. deep-learning reconstruction alters noise texture and therefore many second-order features (see the companion iterative-reconstruction simulation for detail) • Contrast timing and dose: arterial vs. portal-venous phase changes intensity distributions dramatically for hypervascular tumors (e.g., hepatocellular carcinoma) • Voxel resampling: most pipelines resample to isotropic voxels (e.g., 1×1×1 mm) before feature extraction to enable comparison across protocols

These sensitivities motivated the Image Biomarker Standardisation Initiative (IBSI), an international collaboration that defines reference feature calculation formulas and recommends acquisition/preprocessing standards — addressed in detail in the final stage of this simulation.

Segmentation — Defining the Region That Every Feature Depends On

Every radiomic feature is computed exclusively from voxels within a defined tumor mask, making segmentation the single most consequential and most variable step in the entire pipeline. A few millimeters of boundary disagreement between two segmentations can shift texture feature values by 10–30%, particularly for features sensitive to the tumor-normal tissue interface.

• 10–30 min: Manual segmentation time (per 3D tumor volume, experienced reader)
• 0.75–0.85: Inter-observer Dice (manual) (typical for irregular lung/liver tumors)
• 0.85–0.95: nnU-Net Dice (deep learning) (vs. expert ground truth, organ-dependent)
• 10–30%: Feature value sensitivity (change per segmentation variability)

Manual, semi-automatic, and deep-learning segmentation approaches

Manual segmentation: a radiologist or trained analyst draws the tumor boundary on each axial slice using dedicated software (3D Slicer, ITK-SNAP, MIM, syngo.via). This remains the reference standard for many studies but is time-consuming (10–30 minutes per lesion) and subject to inter- and intra-observer variability, particularly for tumors with indistinct or infiltrative margins (e.g., glioblastoma, ground-glass lung nodules).

Semi-automatic segmentation: seeded region-growing, level-set methods, or threshold-based approaches (e.g., a fixed Hounsfield Unit threshold for lung nodules against air background) require a user-placed seed point or bounding box, then algorithmically propagate the boundary. Faster than fully manual delineation and more reproducible, but still requires expert oversight, especially near vessels or adjacent organs.

Deep-learning segmentation: convolutional networks (U-Net, nnU-Net, and more recent transformer-based architectures like Swin-UNETR) trained on large annotated datasets can segment tumors in seconds with Dice similarity coefficients of 0.85–0.95 against expert ground truth for well-represented tumor types (liver, kidney, lung). The nnU-Net framework (Isensee et al., Nature Methods 2021) in particular achieved state-of-the-art results across 23 public segmentation challenges with minimal manual tuning, and is now widely used as the default segmentation engine in radiomics pipelines.

Regardless of method, segmentation quality is typically reported using the Dice similarity coefficient (2×|A∩B| / (|A|+|B|)) against a reference (usually consensus expert) segmentation, with values above 0.80–0.85 generally considered acceptable for downstream radiomics analysis.

Why segmentation variability threatens feature reproducibility

Radiomic features are not uniformly robust to segmentation perturbation. Studies applying controlled boundary erosion/dilation (e.g., ±1–3 mm) to test feature stability find:

• Shape features (volume, sphericity): highly sensitive — small boundary changes directly alter volume and surface-area-derived metrics • First-order statistics (mean, median): relatively robust — dominated by the bulk of interior voxels, less affected by boundary voxels • Texture features (GLCM, GLRLM): moderately sensitive — boundary voxels straddling the tumor-normal interface introduce partial-volume intensity values that distort local co-occurrence patterns, especially in small lesions where boundary voxels are a large fraction of the total

This has motivated "robust radiomics" methodology: many pipelines now explicitly test and report each feature's intraclass correlation coefficient (ICC) across repeated or perturbed segmentations, retaining only features with ICC >0.75–0.90 for downstream model building — directly linking the segmentation stage to the feature-selection stage described later in this simulation.

A 2016 study by Parmar et al. found that of 440 candidate radiomic features, only 30% showed high reproducibility (ICC >0.90) across test-retest CT scans and multiple segmentation methods — underscoring why feature selection and stability filtering are essential rather than optional steps.

First-Order Intensity Statistics and 3D Shape Descriptors

The simplest and most interpretable radiomic features are computed directly from the distribution of voxel intensity values within the tumor mask (first-order/histogram features) and from the geometry of the mask itself (shape/morphology features) — no spatial relationship between neighboring voxels is considered at this stage.

• 18–19: Typical first-order features (per IBSI standard feature set)
• 14: Typical shape features (volume, surface area, sphericity, etc.)
• 0 (irregular)–1 (perfect sphere): Sphericity range (malignant tumors trend lower)
• ρ≈0.4–0.6: Entropy correlation w/ grade (higher entropy, higher tumor grade)

First-order (histogram) features — quantifying the intensity distribution

First-order features summarize the gray-level (Hounsfield Unit, in CT) histogram of all voxels within the tumor mask, ignoring their spatial arrangement:

• Mean, median, minimum, maximum, range: basic central tendency and spread of attenuation values • Standard deviation, variance: overall intensity heterogeneity • Skewness: asymmetry of the intensity distribution (positive skew indicates a tail of high-intensity voxels, e.g., calcification or hemorrhage) • Kurtosis: "peakedness" of the distribution — high kurtosis indicates a sharply peaked histogram with heavy tails, often reflecting a mix of necrotic and viable tumor tissue • Entropy: −Σ p(i)·log₂p(i) over the intensity histogram bins — a measure of randomness/heterogeneity; higher entropy indicates a more heterogeneous, disordered intensity pattern, frequently associated with higher tumor grade and worse prognosis • Energy/uniformity: Σ p(i)² — the inverse concept to entropy; high uniformity indicates a homogeneous tumor • Percentiles (10th, 90th): robust descriptors less sensitive to extreme outlier voxels than min/max

These features are computed after intensity discretization — binning the continuous HU range into a fixed number of gray levels (commonly 16–64 bins, or a fixed bin width of 25 HU) — a critical preprocessing choice governed by IBSI recommendations, since bin count strongly affects entropy and texture feature magnitudes.

Shape and morphology — the geometry of malignancy

Shape features are computed purely from the 3D tumor mask geometry, independent of intensity values, and describe how the tumor is spatially arranged:

• Volume: total voxel count × voxel volume — directly relevant to staging and response assessment (RECIST uses unidimensional diameter, but volumetric measures are more sensitive to subtle change) • Surface area: total mesh surface area of the segmented tumor boundary • Surface-to-volume ratio: higher values indicate a more irregular, spiculated, or lobulated boundary — classically associated with malignancy (vs. the smooth, low surface-to-volume ratio of benign, well-circumscribed lesions) • Sphericity: (36π·V²)^(1/3) / SA — ranges from 0 to 1, with 1 representing a perfect sphere; malignant tumors, particularly those with infiltrative growth patterns, tend toward lower sphericity • Compactness, elongation, flatness: derived from the eigenvalues of the shape's principal component analysis (PCA) covariance matrix, describing whether the tumor is more rod-like, disc-like, or spherical • Maximum 3D diameter: the longest distance between any two points on the tumor surface, closely related to conventional RECIST measurements

Shape features have particular clinical resonance because they formalize long-standing radiological pattern-recognition heuristics — "spiculated margin," "lobulated contour" — into precise, reproducible numbers that can be tracked serially and compared across large cohorts.

Second-Order Texture Features — GLCM, GLRLM, GLSZM, and Filtered Derivatives

Texture features capture the spatial relationships between neighboring voxel intensities — information entirely invisible to first-order histogram statistics. These gray-level matrices are the mathematical heart of radiomics, quantifying concepts like coarseness, contrast, and directional homogeneity that radiologists intuitively perceive but rarely measure numerically.

• ~24: GLCM features (per direction/distance) (contrast, correlation, homogeneity, etc.)
• 8: Wavelet decomposition sub-bands (LLL,LLH,LHL,...,HHH (3D))
• 700–1800: Total features after filtering (original + LoG + wavelet images)
• 13: GLCM matrix directions (3D) (unique angles, often averaged)

Gray-Level Co-occurrence Matrix (GLCM) and Run-Length Matrix (GLRLM)

GLCM (Haralick et al., 1973): a matrix P(i,j) counting how often gray-level i is found adjacent to gray-level j at a specified distance and direction (in 3D, 13 unique directions are typically evaluated and averaged for rotational invariance). From this matrix, classical Haralick features are computed:

• Contrast: Σ(i−j)²·P(i,j) — measures local intensity variation; high contrast indicates sharp, high-frequency texture • Correlation: measures linear dependency of gray levels between neighboring voxels • Homogeneity (inverse difference moment): Σ P(i,j)/(1+|i−j|) — high for images with little local variation • Energy (angular second moment): Σ P(i,j)² — high for orderly, repetitive texture patterns • Entropy: randomness of the co-occurrence distribution itself

GLRLM (Galloway, 1975): counts "runs" — consecutive voxels of the same gray level along a given direction. Features derived include Short Run Emphasis (SRE, weighting toward fine texture), Long Run Emphasis (LRE, weighting toward coarse texture), and Run-Length Non-Uniformity (RLN, measuring how evenly run lengths are distributed).

GLSZM (Gray-Level Size Zone Matrix, Thibault et al. 2009): counts connected zones (not just linear runs) of the same gray level — a rotation-invariant alternative to GLRLM that captures 2D/3D connected regions of homogeneous intensity, useful for describing zone-based heterogeneity (e.g., necrotic pockets vs. diffusely infiltrative tumor).

NGTDM (Neighborhood Gray-Tone Difference Matrix): quantifies the difference between a voxel's gray level and the average of its neighbors, yielding features like Coarseness, Contrast, Busyness, Complexity, and Strength — closely aligned with human visual perception of texture coarseness/fineness.

Filtered/derived images — wavelet and Laplacian-of-Gaussian transforms

To capture texture information at multiple spatial scales and frequency bands — information not visible in the original image alone — radiomics pipelines apply mathematical filters before re-computing the full first-order and texture feature set on each filtered image:

• Laplacian of Gaussian (LoG): a spatial bandpass filter parameterized by σ (sigma, in mm), which enhances edges/blob-like structures at a chosen scale. Fine σ (1–2mm) highlights small textural detail; coarse σ (4–5mm) highlights larger structural patterns. Multiple σ values are often used to build a multi-scale feature bank.

• Wavelet decomposition: a 3D discrete wavelet transform decomposes the image into 8 sub-band images per level, combining high-pass (H) and low-pass (L) filtering along each of the three spatial axes (e.g., LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH). Each sub-band emphasizes different combinations of coarse structure and high-frequency edge/texture information along different axes.

Because the full first-order + GLCM + GLRLM + GLSZM + NGTDM feature battery (typically 100+ features) is recomputed on the original image AND on every LoG-filtered and wavelet sub-band image, a single tumor segmentation commonly yields 700–1,800+ total candidate radiomic features — a dimensionality far exceeding typical cohort sizes (hundreds of patients), which is precisely why the feature selection stage that follows is not optional but essential to avoid catastrophic overfitting.

The Image Biomarker Standardisation Initiative (IBSI, Zwanenburg et al. 2020, Radiology) established consensus reference values for 174 standardized radiomic features computed from digital phantoms, enabling different software implementations (PyRadiomics, MaZda, IBEX, CERR) to be benchmarked and cross-validated — a critical step toward clinical-grade reproducibility.

Feature Selection and Predictive Model Building

With 700–1,800+ candidate features and typically only hundreds of patients per cohort, the curse of dimensionality makes naive model fitting prone to severe overfitting. Feature selection algorithms reduce this space to a compact, non-redundant, and biologically or statistically meaningful signature before a final predictive model is trained.

• 700–1800: Typical starting feature count (from multi-filter extraction)
• 5–20: Typical final signature size (features after selection)
• LASSO, mRMR, RF-importance: Common selection methods (sparsity or redundancy-based)
• 0.70–0.90: Typical validated AUC (for outcome/mutation prediction)

Dimensionality reduction — LASSO, mRMR, and stability filtering

LASSO (Least Absolute Shrinkage and Selection Operator): a regularized regression that adds an L1-penalty term (λ·Σ|β|) to the standard regression loss, driving many feature coefficients exactly to zero. The regularization strength λ is tuned via cross-validation, and the surviving non-zero-coefficient features form the selected signature. LASSO is by far the most commonly used selection method in published radiomics studies because it simultaneously performs selection and produces an interpretable linear model.

mRMR (minimum Redundancy Maximum Relevance): an information-theoretic filter method that ranks features by their mutual information with the outcome (relevance) while penalizing mutual information between selected features themselves (redundancy) — since many texture features derived from the same GLCM or the same wavelet sub-band are highly correlated with one another and provide little independent information.

Stability/reproducibility filtering: as introduced in the segmentation stage, features are often pre-filtered by their intraclass correlation coefficient (ICC) across test-retest scans or segmentation perturbations, retaining only features with ICC >0.75–0.90 before any relevance-based selection is applied — ensuring the final signature is not just predictive in-sample but reproducible across scanners and readers.

A typical pipeline chains these approaches: (1) remove features with ICC <0.75 (reproducibility filter), (2) remove features with pairwise Spearman correlation >0.9 to a higher-ranked feature (redundancy filter), (3) apply LASSO or mRMR for final relevance-based selection, converging on a signature of roughly 5–20 features from the original 700–1,800+ candidates.

Model architectures and validation methodology

Once a compact feature signature is selected, several model classes are commonly applied depending on the clinical task:

• Cox proportional hazards regression: for time-to-event outcomes (overall survival, progression-free survival) — produces a "radiomics score" that stratifies patients into risk groups, often combined into a nomogram with clinical variables (stage, age, histology) • Random forest / gradient boosting (XGBoost): ensemble tree-based classifiers robust to feature scaling and nonlinear relationships, commonly used for binary classification tasks (e.g., benign vs. malignant, mutation-positive vs. -negative) • Support vector machine (SVM): effective in high-dimensional, modest-sample-size settings typical of radiomics, particularly with linear or RBF kernels • Logistic regression with elastic net: combines L1 and L2 regularization, often used when a interpretable, clinically deployable model with a small feature count is prioritized

Rigorous validation requires strict separation of training and test data: an internal train/validation split (often via k-fold or nested cross-validation to jointly tune feature selection and model hyperparameters) followed by evaluation on a genuinely independent external test cohort — ideally from a different institution and scanner — to test generalizability. Studies that only report cross-validated internal performance without external validation are considered at high risk of overoptimistic reported accuracy, a well-documented pitfall in the radiomics literature (Radiomics Quality Score, Lambin et al. 2017, explicitly penalizes lack of external validation).

The Radiomics Quality Score (RQS), a 16-item checklist proposed by Lambin et al. (2017, Nat. Rev. Clin. Oncol.), is widely used to critically appraise published radiomics studies; systematic reviews applying RQS to the literature have repeatedly found median scores below 50% of the maximum, reflecting widespread methodological shortcomings including small sample sizes and absent external validation.

Radiogenomics and Standardization — Linking Imaging Phenotype to Molecular Biology

Radiogenomics extends radiomics one step further: correlating quantitative imaging features with the underlying genomic, transcriptomic, or mutational landscape of the tumor, aiming to non-invasively infer molecular subtype from images alone — a "virtual biopsy" that could guide targeted therapy selection without an additional invasive procedure.

• 0.70–0.85: EGFR-mutant NSCLC imaging AUC (radiomics-predicted mutation status)
• 0.86–0.92: IDH1-mutant glioma prediction AUC (MRI radiogenomics, multiple studies)
• >200: TCIA public imaging-genomic datasets (The Cancer Imaging Archive)
• 174: IBSI standardized feature definitions (consensus reference values published 2020)

Radiogenomics — imaging as a non-invasive proxy for tumor genotype

Radiogenomics studies pair radiomic feature signatures with paired genomic data (typically next-generation sequencing panels or whole-exome/transcriptome sequencing) from the same tumors, using machine learning to identify imaging patterns predictive of specific mutations, gene-expression subtypes, or copy-number alterations. Prominent validated examples include:

• EGFR mutation status in lung adenocarcinoma: EGFR-mutant tumors tend to show ground-glass predominant morphology, air bronchograms, and lower texture heterogeneity on CT compared to EGFR-wildtype tumors; multiple studies report radiomics-based prediction AUCs of 0.70–0.85, relevant because EGFR status determines eligibility for tyrosine kinase inhibitor therapy (osimertinib, erlotinib)

• IDH1 mutation status in glioma: IDH-mutant gliomas (associated with substantially better prognosis) show characteristic MRI texture and diffusion features distinguishable from IDH-wildtype tumors with reported AUCs of 0.86–0.92 in multiple validated cohorts — highly clinically relevant since IDH status is a defining criterion in the WHO 2021 CNS tumor classification

• MGMT promoter methylation in glioblastoma: predictive of response to temozolomide chemotherapy; radiogenomic MRI signatures have shown moderate predictive value (AUC ~0.70–0.80), though less robust than IDH prediction

• Molecular subtypes in breast cancer (luminal A/B, HER2-enriched, triple-negative): dynamic contrast-enhanced MRI radiomics features correlate with receptor status and have been explored as a pre-biopsy triage tool

The biological rationale underlying these correlations is that specific driver mutations alter tumor growth patterns, angiogenesis, cellular density, and necrosis — all of which leave macroscopic imaging signatures detectable by texture and shape analysis, even though the causal chain from single-gene mutation to voxel-level image pattern remains only partially understood mechanistically.

The Cancer Imaging Archive (TCIA), hosted by the NCI, provides more than 200 publicly available, de-identified imaging collections paired with genomic, clinical, and pathological annotations (e.g., TCGA-LUAD, TCGA-GBM), forming the backbone dataset infrastructure for the vast majority of published radiogenomics research.

Standardization, IBSI, and the path to clinical translation

The single greatest barrier to clinical adoption of radiomics and radiogenomics has been reproducibility: different software packages historically computed the "same" named feature (e.g., GLCM contrast) using subtly different mathematical conventions (bin discretization, distance weighting, matrix symmetrization), producing numerically different results from identical images.

The Image Biomarker Standardisation Initiative (IBSI), formalized in a landmark 2020 Radiology publication (Zwanenburg et al.), addressed this by: (1) publishing precise mathematical reference definitions for 174 radiomic features, (2) creating digital reference phantoms with known ground-truth feature values, and (3) benchmarking major software packages (PyRadiomics, MaZda, IBEX, LIFEx, CERR) against these references, achieving consensus agreement for the large majority of tested features.

Beyond software-level standardization, reproducibility also requires: consistent voxel resampling/interpolation, standardized intensity discretization (fixed bin number vs. fixed bin width — IBSI recommends reporting both), harmonization of acquisition protocols across scanners (or statistical harmonization methods such as ComBat, borrowed from genomics batch-effect correction, increasingly applied to multi-center radiomics datasets), and mandatory external validation cohorts.

With IBSI compliance now a de facto requirement for high-quality publication and with major professional societies (RSNA, EORTC) issuing radiomics reporting guidelines, the field is transitioning from a purely exploratory research tool toward regulatory-grade, clinically deployable quantitative imaging biomarkers — several FDA-cleared and CE-marked radiomics-based software products already exist for specific applications including lung nodule risk stratification and prostate cancer characterization.