Radiomic Features in Multi-parametric MRI

• Radiomic features in multi-parametric MRI are quantitative descriptors derived from anatomical and functional sequences that capture intensity, texture, and shape information within lesions.
• They enhance noninvasive tissue characterization, tumor grading, and biomarker prediction by integrating multi-sequence correlations and advanced statistical methods.
• Robust feature extraction depends on standardized acquisition, preprocessing, segmentation accuracy, and harmonization strategies to improve reproducibility and clinical impact.

Radiomic features in multi-parametric MRI (mpMRI) comprise a high-dimensional set of quantitative descriptors computed from anatomical and functional MRI sequences. Extracted from expertly or algorithmically segmented regions of interest (ROIs), these features encode information about intensity distributions, textural patterns, and geometric properties, as well as multi-sequence correlations. Their systematic use has emerged as a cornerstone for noninvasive tissue characterization, tumor grading, response assessment, and prediction of molecular biomarkers across a spectrum of malignancies, notably in brain, prostate, and breast oncology. The robustness, biological correspondence, and clinical utility of radiomic features hinge critically on acquisition protocol, intensity normalization, segmentation accuracy, feature selection, and harmonization strategies.

1. Core Radiomic Feature Families and Definitions

Radiomic features encompass several families, each encoding a specific aspect of lesion morphology or texture. Standard categories, following the IBSI and PyRadiomics frameworks, include:

• First-order statistics: Histogram-based descriptors within the ROI, such as mean, variance, skewness, kurtosis, energy, and entropy.
• Shape/Size metrics: Volume, surface area, sphericity, compactness, axis lengths, and surface-area-to-volume ratio. For instance, the least axis length (LAL) in 3D is $\sqrt{\lambda_{\min}}$, where $\lambda_{\min}$ is the smallest eigenvalue of the lesion mask's covariance matrix (Salmanpour et al., 14 Dec 2024).
• Gray-Level Co-occurrence Matrices (GLCM): Pairwise spatial statistics, e.g., contrast $\sum_{i,j}(i-j)^2 P(i,j)$, correlation, homogeneity, and energy (Schwier et al., 2018, Cui et al., 2019, Salmanpour et al., 14 Dec 2024).
• Gray-Level Run Length Matrix (GLRLM): Captures consecutive runs, features like run entropy for ADC images indicate tissue heterogeneity (Salmanpour et al., 14 Dec 2024).
• Gray-Level Size Zone Matrix (GLSZM): Measures connected zones with the same gray level.
• Gray-Level Dependence Matrix (GLDM): Quantifies the dependence of a center voxel on its neighbors.
• Neighbouring Gray-Tone Difference Matrix (NGTDM): Measures differences in intensity between a center voxel and its neighborhood.

High-order and multi-parametric features extend these concepts to joint-probabilities or wavelet-decomposed images, e.g., tissue-signature probability and co-occurrence matrices for multiple MRI contrasts (Parekh et al., 2018, Parekh et al., 2019).

2. Standard mpMRI Acquisition, Preprocessing, and Normalization

Accurate computation of radiomic features requires careful protocol standardization:

• Multi-contrast acquisition: Typical mpMRI includes T1, T1-contrast, T2, FLAIR, DWI-derived ADC, and dynamic contrast-enhanced (DCE) series (Li et al., 2021, Chen et al., 2023, Feng et al., 15 Oct 2025, Salmanpour et al., 14 Dec 2024).
• Image preprocessing: Sequences undergo spatial co-registration (rigid, affine, or nonrigid), slice resampling, and intensity normalization. Common schemes include z-score normalization within mask or whole volume, min–max normalization, and histogram matching. Bias-field (N4) correction is standard for compensation of scanner inhomogeneities (Schwarzhans et al., 3 Jun 2024, Schwier et al., 2018).
• Segmentation: Manual or algorithmic delineation of lesions, with consensus review when available, is critical for feature stability (Salmanpour et al., 14 Dec 2024, Li et al., 2021).
• Discretization: Quantization of intensity values (fixed bin width or bin count) affects texture metrics' reproducibility. Bin widths of 15–20 or producing 32–64 gray levels per ROI are typical (Schwier et al., 2018, Li et al., 2021).

Normalization pipelines profoundly affect feature robustness. For example, in breast MRI, combining N4 bias correction with piecewise linear histogram equalization (PLHE) or z-score normalization restricted to the tissue of interest yields the most stable and predictive features (median ICC for GLCM/GLRLM/GLDM > 0.8 under linear pipelines, and key “nonuniformity” features ICC > 0.9 across all methods) (Schwarzhans et al., 3 Jun 2024, Schwier et al., 2018).

3. Multi-parametric Feature Extraction and High-order Methods

Conventional radiomics compute features independently per sequence and concatenate results. Advanced multiparametric frameworks produce integrated representations:

• Joint feature matrices: For N co-registered MRI volumes, the tissue-signature probability matrix (TSPM) encodes the joint distribution:

$$P_{TSPM}(i_1,\ldots,i_N) = \frac{M(i_1,\ldots,i_N)}{\sum M}$$

where $M$ is the count of voxels $p$ with signature $S_p = (i_1,\ldots,i_N)$ (Parekh et al., 2018, Parekh et al., 2019).

• Multiparametric Texture: The tissue-signature co-occurrence (TSCM) and complex interaction network features extend GLCM/GLRLM concepts to high-dimensional spaces, capturing inter-sequence relationships inaccessible to single-image radiomics (Parekh et al., 2018).
• Spherical/decomposition approaches: Novel geometries, such as spherical radiomics, analyze tumor surfaces by extracting features along concentric radial shells, improving correlation with underlying molecular heterogeneity compared to the traditional Cartesian approach (Feng et al., 15 Oct 2025).
• Wavelet decomposition: Feature extraction on multiple spatial-frequency sub-bands (e.g., 8 wavelet bands per image) increases sensitivity to multiscale texture heterogeneity (Bobholz et al., 2019).
• Deep/discovery radiomics: Data-driven feature learning (deep convolutional “sequencer” architectures) can outperform hand-crafted feature sets for subtype classification and lesion detection, as in prostate mpMRI (Chung et al., 2015).

4. Robustness, Generalizability, and Harmonization

Rigorous assessment of radiomic feature repeatability and reproducibility includes:

• Test–retest repeatability: Measured via Intraclass Correlation Coefficient (ICC), values above 0.85 are considered excellent but are highly sensitive to binning, filtering, and normalization scheme (Schwier et al., 2018). Parameters such as spatial filtering (LoG, wavelets), 2D vs. 3D computation, and ROI delineation method all significantly affect ICC. Some configuration/feature combinations yield ICC < 0, underscoring the need for dataset-specific assessment.
• Between-center harmonization: Batch effects from hardware or protocol differences can be reduced via statistical methods such as ComBat harmonization, which improves cross-center median ICCs (e.g., from 0.62 to 0.81) and enhances external cohort prediction (Li et al., 2021).
• Feature selection and reporting: Combining univariate (Wilcoxon, ANOVA F-test), filter (mutual information, variance threshold), and embedded methods (LASSO, logistic regression coefficients) isolates compact, interpretable, and predictive feature sets (Salmanpour et al., 14 Dec 2024, Cui et al., 2019). Reporting all processing and selection parameters is essential for reproducibility.
• Cross-validation and external validation: All robust models validate extraction and model performance through external or hold-out test sets, using AUC, sensitivity, specificity, and statistical hypothesis testing (e.g., DeLong, paired t-test) (Schwarzhans et al., 3 Jun 2024, Li et al., 2021, Cui et al., 2019).

5. Biological Relevance and Clinical Impact

Radiomic features correlate with biological ground truth at multiple levels:

• Correlation with histology: In brain cancer, first-order radiomic features (e.g., FLAIR FO energy, T1+C FO variance) show highest correspondence (Spearman’s ρ ≈ 0.45–0.60) with their histomic analogues, are most stable across scanner confounds, and reflect underlying tissue architecture (Bobholz et al., 2019).
• Molecular biomarker prediction: Features derived on spherical radial shells identify molecular stratification in glioblastoma (MGMT, EGFR, PTEN), with GLCM features dominating model importance (Feng et al., 15 Oct 2025).
• Clinical prediction and workflow: Automated extraction (following expert segmentation) enables predictive models for tumor grade, treatment response (AUC up to 0.95 for mpRadiomics in brain tumors (Parekh et al., 2019)), and decision support in ambiguous PI-RADS 3 prostate lesions, with up to 30% fewer unnecessary biopsies (Li et al., 2021).
• Compact, interpretable signatures: A handful of features—intensity upper tail (T2WI 90th percentile), variance, LAL, SAVR, GLRLM entropy—jointly achieve high cross-validated accuracy for prostate cancer risk; these are directly linked to PI-RADS semantics (Salmanpour et al., 14 Dec 2024).

6. Limitations, Best Practices, and Prospects

Radiomic workflows are sensitive to all preprocessing and extraction details, and single-feature/threshold generalizability is limited (Schwier et al., 2018, Schwarzhans et al., 3 Jun 2024). Key recommendations include:

• Always include bias-field correction and report normalization equations and discretization parameters (Schwarzhans et al., 3 Jun 2024, Schwier et al., 2018).
• Preprocessing choices must be optimized per-dataset; test–retest ICC and cohort-based harmonization must precede claims of biomarker robustness.
• Limit feature sets using integrated univariate/multivariate selection, with preference for features robust to pipeline variation and interpretable from a biological/clinical standpoint (Salmanpour et al., 14 Dec 2024).
• Avoid over-reliance on unidimensional metrics (e.g., ICC) and always report performance benchmarks on external data (Schwarzhans et al., 3 Jun 2024, Li et al., 2021).
• Future directions include fully automated pipelines (deep segmentation, discovery radiomics), leveraging hybrid deep–radiomic models, expanding to other organs, and multi-omic integration (radiogenomics, pathomics) (Parekh et al., 2018, Chung et al., 2015).

In sum, radiomic features in mpMRI offer a powerful, quantitative link between image phenotype and tissue biology, but their deployment in research and clinical workflows requires standardized acquisition, rigorous preprocessing, robust validation, and explicit, transparent reporting (Schwarzhans et al., 3 Jun 2024, Schwier et al., 2018, Salmanpour et al., 14 Dec 2024).