Z-Score Normalization

The z-score is calculated by subtracting the population mean (μ) from the raw value (x) and dividing the result by the standard deviation (σ).

• Formula: z = (x - μ) / σ
• Result: A distribution with a mean of 0 and a standard deviation of 1.
• Unitless: The transformation strips the original physical units (e.g., Hounsfield Units), converting them to standard deviation units.

This technique centers the data cloud around the origin and scales the axes equally.

• Centering: Shifts the data so the arithmetic mean sits exactly at zero.
• Scaling: Expands or contracts the variance so the spread is uniform across features.
• Shape Preservation: Unlike min-max scaling, z-score normalization preserves the shape of the original distribution, including outliers and skewness.

Z-score normalization does not bound data to a fixed range, making it sensitive to extreme values.

• Unbounded Range: Values can theoretically range from negative to positive infinity.
• Outlier Detection: A z-score magnitude greater than 3 is often considered a statistical outlier in a normal distribution.
• Robust Alternative: For highly skewed radiomic data, consider Robust Scaling (using median and interquartile range) to mitigate outlier influence.

Many machine learning algorithms explicitly assume standardized input features for optimal convergence.

• Distance-Based Models: Essential for k-Nearest Neighbors (k-NN) and Support Vector Machines (SVM) to prevent high-magnitude features from dominating distance calculations.
• Gradient Descent: Required for logistic regression and neural networks to ensure smooth, efficient convergence without oscillation.
• Principal Component Analysis (PCA): Mandatory to ensure the first principal component captures the direction of maximum variance, not just the variable with the largest scale.

In radiomics, the reference statistics (μ, σ) can be calculated from different populations, affecting generalizability.

• Global Normalization: μ and σ are calculated from the entire training dataset. This is standard practice.
• Batch Normalization: μ and σ are calculated from a mini-batch during neural network training, acting as a regularizer.
• ComBat Harmonization: A specialized extension that adjusts for batch effects across different imaging scanners before applying standard z-score normalization.

The process is fully reversible, allowing model outputs to be mapped back to the original physical units for clinical interpretation.

• Formula: x = (z * σ) + μ
• Clinical Relevance: A predicted standardized biomarker can be converted back to Hounsfield Units or SUV values.
• Reproducibility: Requires storing the exact μ and σ values from the training set to apply to future unseen patient scans.