Robust Feature Selection

The primary statistical tool for quantifying test-retest reproducibility. ICC measures the consistency of a feature's value when the same subject is scanned twice within a short interval.

• ICC > 0.75: Generally considered good reliability
• ICC > 0.90: Excellent stability, suitable for clinical biomarkers
• Features below a pre-defined ICC cutoff are discarded before model training

This ensures the final radiomic signature is driven by true biological signal rather than scanner noise.

CCC evaluates both precision and accuracy simultaneously, measuring how closely feature values from different observers or scanners fall on the 45-degree agreement line.

• Combines Pearson correlation with a bias correction factor
• More stringent than ICC alone; penalizes systematic offsets
• Essential for multi-center trials where inter-scanner variability is high

Features with high CCC are considered robust to both random error and systematic bias.

Features with near-zero variance across a patient population provide no discriminatory power and are removed prior to modeling.

• Coefficient of Variation (CV): Ratio of standard deviation to mean; low CV indicates a stagnant feature
• Biological Plausibility Check: Features that violate known anatomical constraints (e.g., negative volume) are flagged
• Removes redundant features that would only add noise to the model

Highly correlated feature groups are collapsed into single representative variables to avoid multicollinearity.

• Spearman's rank correlation matrix identifies clusters of redundant features
• From each cluster, the feature with the highest ICC is retained as the cluster representative
• Reduces feature space by 40-60% while preserving information content

This step is critical before applying LASSO or other regularized regression techniques.

Features are stress-tested against inter-observer segmentation uncertainty by perturbing the region of interest boundary.

• Erosion/dilation simulations: The ROI is systematically expanded and contracted by 1-3 voxels
• Features whose values fluctuate dramatically with minor boundary changes are deemed unstable
• Stable features show consistent values regardless of marginal segmentation differences

This directly addresses a primary source of irreproducibility in clinical radiomics.

Unwanted technical variation from scanner manufacturer, acquisition protocol, or reconstruction kernel is identified and mitigated before feature selection.

• Principal Component Analysis (PCA) visualizations colored by scanner type reveal batch effects
• ComBat harmonization adjusts feature values to remove systematic technical biases
• Only post-harmonization stable features are advanced to model building

This ensures a signature developed at one institution can be deployed at another.

In oncology, robust feature selection isolates imaging phenotypes that are stable across CT scanners and segmentation variability. By filtering out non-reproducible texture features, models predicting overall survival or distant metastasis in lung and head-and-neck cancers achieve higher external validation concordance indices (C-index).

• Eliminates features with Intraclass Correlation Coefficient (ICC) < 0.75
• Reduces dimensionality from thousands of features to a parsimonious signature of 5-15 stable biomarkers
• Ensures prognostic models are vendor-agnostic and portable across hospital networks

Delta-radiomics workflows depend on robust feature selection to identify which textural changes between pre- and post-treatment scans reflect true biological response rather than measurement noise. Features with high test-retest repeatability are selected to quantify early response to immunotherapy and chemoradiation.

• Tracks temporal heterogeneity changes in non-small cell lung cancer
• Uses Concordance Correlation Coefficient (CCC) to filter stable delta features
• Predicts pathological complete response (pCR) in neoadjuvant trials

Non-invasive prediction of IDH mutation and 1p/19q co-deletion status relies on robust radiomic features extracted from multi-parametric MRI. Feature selection via LASSO regularization with bootstrap resampling identifies texture and shape features stable across inter-rater segmentation variability.

• Integrates features from T1, T1-contrast, T2, and FLAIR sequences
• Retains only features with Cohen's κ > 0.8 across multiple radiologist segmentations
• Achieves AUC > 0.85 in external validation cohorts for molecular subtyping

Robust feature selection identifies imaging surrogates for tumor-infiltrating lymphocyte (TIL) density and PD-L1 expression. By applying Minimum Redundancy Maximum Relevance (mRMR) with stability analysis, researchers isolate texture features that correlate with the tumor immune microenvironment independent of scanner protocol.

• Links GLCM entropy and GLSZM zone percentage to immune infiltration
• Uses perturbation-based stability to validate feature importance rankings
• Enables non-invasive patient stratification for checkpoint inhibitor eligibility

Prospective clinical trials require radiomic signatures that generalize across heterogeneous acquisition protocols. Robust feature selection employs ComBat harmonization followed by stability filtering to identify features invariant to scanner manufacturer, tube current, and reconstruction kernel differences.

• Applies nested cross-validation to prevent information leakage during selection
• Retains features with Coefficient of Variation (CV) < 15% in phantom studies
• Validates signatures in federated learning settings without centralizing imaging data

Beyond oncology, robust feature selection identifies stable plaque texture features from coronary CT angiography that predict major adverse cardiac events. Features are filtered for inter-observer reproducibility in delineating arterial wall boundaries.

• Extracts pericoronary fat attenuation and plaque heterogeneity metrics
• Uses Bland-Altman analysis to quantify inter-observer agreement limits
• Builds radiomic signatures additive to traditional Agatston calcium scoring

The computational process of removing unwanted technical variability from radiomic features caused by differences in scanner manufacturers, acquisition protocols, or reconstruction kernels. Without harmonization, a feature selection algorithm may discard biologically meaningful features simply because they appear unstable across heterogeneous imaging batches.

• ComBat Harmonization: Adapts a statistical batch-effect correction method from genomics to standardize feature values across multiple imaging centers.
• Goal: Retain biological signal while eliminating scanner-induced noise.

The assessment of radiomic feature stability when imaging is repeated on the same subject under identical conditions within a short interval. Features with high test-retest variability are inherently unreliable and must be excluded before model building.

• Measured via Intraclass Correlation Coefficient (ICC) or Concordance Correlation Coefficient (CCC).
• A common threshold for retaining a feature is ICC > 0.75 or CCC > 0.85.

A mathematical process for reducing the number of random variables under consideration, essential when the number of extracted radiomic features far exceeds the number of patient samples. Robust feature selection is one class of dimensionality reduction, but it is often paired with unsupervised transformation techniques.

• Principal Component Analysis (PCA): Converts correlated features into linearly uncorrelated principal components.
• t-SNE / UMAP: Non-linear techniques used primarily for visualization of high-dimensional feature spaces.

The process of converting continuous image intensity values into a finite number of discrete bins, a critical pre-processing step required before calculating texture matrices like GLCM or GLRLM. The choice of bin width directly impacts the robustness and reproducibility of the resulting texture features.

• Fixed Bin Number (FBN): Divides the intensity range into a constant number of bins.
• Fixed Bin Width (FBW): Uses a constant bin size in Hounsfield Units, generally preferred for cross-scanner robustness.

The interpolation of medical image data to create isotropic voxels—cubes with equal dimensions in all three spatial axes. This ensures that spatial relationship measurements, such as those captured by GLCM offsets or shape features, are directionally consistent and not biased by the original anisotropic acquisition resolution.

• Common target resolution: 1x1x1 mm³ isotropic.