Minimum Redundancy Maximum Relevance (mRMR)

The algorithm first identifies features with the highest statistical dependency on the target outcome variable. Relevance is typically quantified using mutual information, which measures the amount of information one variable provides about another. For a target class c and feature set S, the max-relevance condition maximizes the average mutual information between individual features x_i and the target:

• Goal: Select features that are most predictive of the clinical endpoint
• Metric: max D(S,c) where D = (1/|S|) Σ I(x_i; c)
• Example: In tumor grading, SUVmax and entropy may show high individual relevance to malignancy status

Features selected solely on relevance often exhibit high inter-correlation, leading to redundant information and inflated dimensionality. The min-redundancy condition penalizes features that share high mutual information with already-selected features:

• Goal: Eliminate features that carry overlapping information
• Metric: min R(S) where R = (1/|S|²) Σ I(x_i; x_j)
• Impact: Prevents multicollinearity and reduces overfitting in downstream radiomic models
• Example: Volume and maximum diameter are often redundant; mRMR retains only one

mRMR employs a greedy forward selection approach rather than exhaustive search, which would be computationally intractable for high-dimensional radiomic feature sets. The algorithm iteratively adds one feature at a time:

• First feature: Selected purely on maximum relevance to the target
• Subsequent features: Chosen to maximize the mRMR score = relevance minus redundancy
• Optimization: max Φ(D,R) where Φ = D - R (MID scheme) or Φ = D / R (MIQ scheme)
• Stopping criterion: Predefined number of features k or threshold on incremental gain

The algorithm offers two primary operational schemes for combining relevance and redundancy into a single objective function:

• MID (Mutual Information Difference): Φ = D - R — subtracts redundancy from relevance directly, favoring features with high net information gain
• MIQ (Mutual Information Quotient): Φ = D / R — divides relevance by redundancy, emphasizing features with high relevance-to-redundancy ratios
• Selection guidance: MID tends to select more features with high absolute relevance; MIQ is more aggressive at suppressing redundancy
• Practical note: Both schemes often yield similar top-ranked features but may diverge in mid-ranked selections

mRMR is particularly well-suited to radiomic biomarker discovery where feature sets are vast (hundreds to thousands of features) and sample sizes are limited:

• Input: PyRadiomics-extracted feature matrices containing shape, first-order, and texture features
• Preprocessing: Features are typically z-score normalized and discretized before mRMR application
• Integration: Often paired with LASSO or SVM-RFE for final model building after mRMR reduces the feature pool
• Validation: Selected features must demonstrate stability via Intraclass Correlation Coefficient (ICC) across test-retest scans
• Clinical impact: Enables parsimonious radiomic signatures with fewer features than patients, satisfying statistical power requirements

mRMR occupies a specific niche in the feature selection landscape, balancing filter efficiency with multivariate awareness:

• vs. Univariate filters (t-test, chi-squared): mRMR accounts for feature interdependencies that univariate methods ignore
• vs. LASSO: LASSO performs embedded selection during model training; mRMR is model-agnostic and can precede any classifier
• vs. Principal Component Analysis (PCA): PCA creates latent variables, losing interpretability; mRMR preserves original feature identities for clinical explainability
• vs. Recursive Feature Elimination (RFE): RFE is wrapper-based and computationally heavier; mRMR is faster for initial high-dimensional screening
• Limitation: Greedy search may miss optimal combinations that require joint evaluation of feature subsets

A modeling error where a statistical model captures noise and random fluctuations in the training data rather than the true underlying signal. In high-dimensional radiomics, where thousands of features are extracted from a small patient cohort, overfitting is a critical risk. mRMR mitigates this by selecting a compact, non-redundant feature subset, reducing the curse of dimensionality and improving the model's ability to generalize to unseen scans from different institutions.

A composite biomarker consisting of a selected panel of quantitative imaging features combined via a mathematical model to predict a specific clinical endpoint, such as overall survival or treatment response. mRMR is a foundational step in signature construction, ensuring the chosen features are maximally informative about the target outcome while being minimally correlated with each other. A robust signature avoids collinearity, which can destabilize model coefficients and reduce interpretability.

A statistical metric used to assess the test-retest reproducibility and inter-observer reliability of radiomic feature measurements. Before applying mRMR, features with low ICC (typically < 0.75) are often filtered out, as they represent noise from segmentation variability rather than true biological signal. mRMR is then applied to the remaining robust features, ensuring the final signature is built on a foundation of stable, reproducible measurements that can be trusted across different radiologists and scanning sessions.

Techniques applied to mitigate systematic technical variation introduced by non-biological factors such as scanner manufacturer, acquisition protocol, or reconstruction kernel. ComBat harmonization is a widely adopted method adapted from genomics. If batch effects are not corrected before feature selection, mRMR may inadvertently select features that discriminate between scanner types rather than clinical outcomes, leading to a signature that fails when deployed across different hospital systems.

A feature scaling technique that standardizes radiomic feature values by centering them to a mean of zero and scaling to a standard deviation of one. This preprocessing step is essential before applying mRMR because the mutual information calculation can be sensitive to the scale of continuous variables. Normalization ensures that features measured in different units—such as tumor volume in cubic millimeters and entropy in dimensionless bits—contribute equally to the relevance and redundancy computations.