Radiomics Biomarkers in Medical Imaging

• Radiomics biomarkers are quantitative imaging descriptors that capture tissue morphology, intensity distributions, and texture patterns for non-invasive tumor characterization.
• They utilize standardized mathematical foundations and preprocessing protocols (e.g., first-order, shape, and texture features) as defined by IBSI for reliable feature extraction.
• These biomarkers are widely applied in oncology, lung cancer, and neurodegeneration to improve diagnosis, prognosis, and therapeutic response assessment.

Radiomics biomarkers are quantitative descriptors derived from medical images, designed to capture tissue morphology, intensity distributions, and higher-order texture patterns not apparent to human readers. These biomarkers serve as non-invasive signatures for tumor phenotyping, diagnosis, prognosis, and therapeutic response assessment across a variety of modalities and disease settings.

1. Mathematical Foundations and Feature Classes

Radiomics features quantify imaging data using standardized mathematical definitions, most commonly following the Image Biomarker Standardization Initiative (IBSI) guidelines. Major feature classes are:

First-Order Statistics: Histogram-based measures such as mean intensity ($\mu = \frac{1}{N}\sum_{i=1}^N x_i$), variance ($\sigma^2 = \frac{1}{N}\sum_{i=1}^N (x_i-\mu)^2$), skewness, kurtosis, percentiles, energy, and entropy ($H = -\sum_{k} p_k \log_2 p_k$) characterize voxel intensity distributions within a region of interest (ROI) (Jouzdani et al., 31 Dec 2025).

Shape Descriptors: Volumetric and geometric features including volume ($V = \sum_{v \in ROI}\mathrm{voxel\_size}$), surface area, sphericity ($\Psi = \frac{\pi^{1/3}(6V)^{2/3}}{A}$), compactness, and principal axes lengths encode 3D lesion morphology (Jouzdani et al., 31 Dec 2025).

Texture Features: Quantify spatial patterns of gray-level variations. Key families include:

• Gray-Level Co-Occurrence Matrix (GLCM): Measures such as contrast ($\mathrm{Contrast} = \sum_{i,j}(i-j)^2 P(i,j)$), correlation, energy, homogeneity.
• Gray-Level Run-Length Matrix (GLRLM): Short-run and long-run emphases, run-length non-uniformity.
• Gray-Level Size-Zone Matrix (GLSZM): Zone size non-uniformity, small/large zone emphasis.
• Gray-Level Dependence Matrix (GLDM): Small dependence emphasis, non-uniformity.
• Neighborhood Gray-Tone Difference Matrix (NGTDM): Coarseness, contrast, busyness.

All formulations are rigorously defined per IBSI (Jouzdani et al., 31 Dec 2025, Lei et al., 2020).

Filter-Based and Multi-Parametric Extensions: Features can be extracted from images pre-processed by convolutional filters (e.g., Laplacian of Gaussian, wavelets) to accentuate specific spatial structures, with strictly defined mathematical kernels and parameter reporting per IBSI-2 (Depeursinge et al., 2020).

2. Extraction, Preprocessing, and Standardization

A robust radiomics biomarker workflow requires uniform image preprocessing and feature computation:

• Image Preprocessing: Common practices include bias-field correction, resampling to isotropic voxels, and intensity normalization (e.g., z-score normalization: $I'(x) = [I(x) - \mu_{ROI}]/\sigma_{ROI}$) (Li et al., 2021).
• Segmentation: ROIs are defined via manual, semi-automated, or deep learning-based segmentation; consistency in methodology is critical (e.g., nnU-Net for airway or liver segmentation) (Mesbah et al., 13 Jun 2025, Hinterberger et al., 27 Oct 2025).
• Feature Extraction Tools: Open-source libraries such as PyRadiomics, PySERA, and RPTK are commonly used, with compliance to IBSI definitions ensuring reproducibility (Salmanpour et al., 20 Nov 2025, Jouzdani et al., 31 Dec 2025).
• Standardization: Direct benchmarking with IBSI digital phantoms is necessary to validate feature outputs. Non-morphological features typically achieve sub-$10^{-6}$ relative differences across compliant software; shape features exhibit greater algorithmic variability (Lei et al., 2020, Salmanpour et al., 20 Nov 2025).

3. Construction and Validation of Radiomics Biomarkers

Radiomics biomarkers are derived through feature selection and modeling:

• Feature Selection: Dimensionality reduction is necessary to prevent overfitting. Typical methods include $\ell_1$-penalized logistic regression (LASSO), univariate statistics (ANOVA, Mann–Whitney U), recursive feature elimination, and tree-based importance rankings (Li et al., 2021, Jouzdani et al., 31 Dec 2025, Kozák, 2024).
• Modeling: Biomarkers are often constructed as sparse linear combinations of selected features or as higher-dimensional signatures:

$$\mathrm{Biomarker\,Score} = \sum_{k=1}^n w_k f_k + b$$

where $w_k$ are learned feature weights, $f_k$ are selected features, and $b$ is the intercept (Li et al., 2021).

• Validation: Multi-center validation frameworks and stratified cross-validation are employed to test biomarker generalizability and robustness to site/scanner variability (Li et al., 2021, Hinterberger et al., 27 Oct 2025). Performance metrics include area under the ROC curve (AUC), sensitivity, specificity, concordance correlation coefficient (CCC), and survival modeling metrics (Chen et al., 2021).

4. Applications Across Modalities and Disease Domains

Radiomics biomarkers are applied in diverse clinical and research contexts:

• Oncology: PI-RADS 3 prostate lesions, non-invasive IDH genotype prediction in glioblastoma, colorectal neoplasia from liver CT, and molecular marker prediction in GBM using spherical radiomics (Li et al., 2021, Kozák, 2024, Feng et al., 15 Oct 2025, Hinterberger et al., 27 Oct 2025).
• Lung Cancer: Mapping radiomics features to Lung-RADS descriptors for interpretable screening (Jouzdani et al., 31 Dec 2025), survival prediction in lung fibrosis using airway-centered radiomics (Mesbah et al., 13 Jun 2025), and hybrid biomarkers combining radiomics with radiologist-provided semantic scores (Mehta et al., 2020, Brocki et al., 2023).
• Neurodegeneration: Shape-based radiomics outperform domain-transferred deep learning features in early Alzheimer's diagnosis from brain MRI (Nielsen et al., 2024).
• Multiple Myeloma: Prognostic stratification via global bone volume/density metrics outperformed local lesion texture features (Schenonea et al., 2020).
• Multiparametric Radiomics: Fusing multiple MRI contrasts and high-order co-occurrence matrices yields high accuracy in brain tumor grading and therapy response (Parekh et al., 2019).
• Functional/Enriched Biomarkers: Integrating classical structural and functional radiomics (e.g., enhancement pattern mapping quantiles smoothed with quantlet bases and Bayesian tensor regression) has advanced risk stratification in hepatic malignancy (Reinhardt et al., 6 Mar 2026).

5. Methodological Innovations and Robustness

Several methodological advances strengthen the clinical readiness of radiomics biomarkers:

• Robustness to Acquisition Variation: Cycle-consistent GANs for denoising low-dose CTs improve reproducibility (CCC from 0.68 to 0.94) and downstream prognostic accuracy (AUC improvements of ~0.05–0.07) (Chen et al., 2021).
• Multi-Flavour Feature Fusion: The Tensor Radiomics paradigm fuses features computed under diverse parameterizations (bin width, segmentation, filters, fusion methods), represented as a three-way tensor $X \in \mathbb{R}^{N \times F \times V}$, where $V$ indexes "flavours" (Rahmim et al., 2022). "TR-Net" fuses multi-flavour features in deep architecture, improving accuracy and reproducibility in multi-task studies (e.g., test–retest ICC $>$90% for the majority of MR features in GBM) (Rahmim et al., 2022).
• Spherical Representation: Mapping tumor volumes to radial shell surfaces improves the detection of biologically relevant gradients and molecular correlates over conventional Cartesian strategies (Feng et al., 15 Oct 2025).
• Automated ML Workflow Optimization: The WORC framework automates all workflow decisions from preprocessing and feature extraction to model selection via AutoML, validated across twelve clinical outcomes, demonstrating improved or comparable AUCs relative to expert observers (Starmans et al., 2021).
• Interpretable Biomarkers: Combining radiomics features with concept bottleneck models (CBM) and SHAP analysis connects quantitative descriptors with clinician-understood semantics, enhancing interpretability and trust (Brocki et al., 2023, Jouzdani et al., 31 Dec 2025).

6. Reproducibility, Standardization, and Best Practices

Reproducibility across centers and software is paramount:

• Standardization: Features and workflows must strictly adhere to IBSI mathematical definitions, including discretization conventions and mesh algorithms for shape (Lei et al., 2020, Depeursinge et al., 2020, Salmanpour et al., 20 Nov 2025).
• Benchmarking: Software must attain sub-percent-level agreement on phantoms for non-shape features; greater discrepancies in shape features must be reported and justified (Lei et al., 2020, Salmanpour et al., 20 Nov 2025).
• Documentation: Complete reporting of all preprocessing, feature configurations, and software versions is mandatory for reproducibility (Lei et al., 2020).
• Automated Libraries: PySERA achieves $>94\%$ IBSI reproducibility and supports both handcrafted and deep radiomics, establishing a robust, standardized computational platform (Salmanpour et al., 20 Nov 2025).

7. Biological Interpretability and Clinical Translation

Radiomics biomarkers provide quantifiable proxies for tissue pathology:

• Pathophysiological Correlates: Texture heterogeneity metrics (e.g., GLCM entropy, joint entropy) capture tumor microenvironment complexity, necrosis, and invasiveness. Shape measures (sphericity, compactness) align with criteria for malignancy risk (Jouzdani et al., 31 Dec 2025, Kozák, 2024).
• Clinical Decision Support: Biomarker cutpoints derived from ROC analysis (e.g., Youden index) achieve high sensitivity and specificity in risk stratification (Li et al., 2021, Hinterberger et al., 27 Oct 2025).
• Interpretability: Mapping radiomics features to standard lexica (Lung-RADS, PI-RADS) and using interpretable models strengthens clinical adoption and trust (Jouzdani et al., 31 Dec 2025, Brocki et al., 2023).

In sum, radiomics biomarkers constitute a rigorous, standardized, and versatile approach to quantitative imaging phenotyping, with established methodology for robust feature extraction, model construction, validation, and clinical translation. Current challenges and research directions relate to further improving reproducibility, multi-center generalizability, functional imaging integration, and biological interpretability (Li et al., 2021, Chen et al., 2021, Reinhardt et al., 6 Mar 2026).