Hounsfield Unit (HU) Rescaling

HU rescaling applies a linear transformation to raw attenuation coefficients (μ) using the formula:

HU = 1000 × (μ - μ_water) / μ_water

• μ_water is the attenuation coefficient of distilled water at standard temperature and pressure
• Air is fixed at exactly -1000 HU
• Water is fixed at exactly 0 HU
• This linear relationship ensures that a 1% change in attenuation relative to water corresponds to a 10 HU shift
• The transformation preserves the relative ordering of tissue densities while normalizing the absolute scale

Different biological tissues occupy characteristic ranges on the Hounsfield scale, enabling automated tissue classification:

• Cortical bone: +400 to +1000 HU (dense calcium hydroxyapatite)
• Soft tissue: +20 to +100 HU (muscle, liver, spleen)
• Water: 0 HU (CSF, simple cysts)
• Fat: -50 to -100 HU (adipose tissue, lipid-rich lesions)
• Lung parenchyma: -700 to -900 HU (air-filled alveoli)

These ranges form the basis for threshold-based segmentation in radiomics pipelines.

Rescaling accuracy depends on regular scanner calibration using standardized phantoms:

• Water phantoms verify the 0 HU reference point
• Multi-density inserts (acrylic, bone-equivalent) validate linearity across the scale
• Daily air calibration checks enforce the -1000 HU anchor
• AAPM guidelines specify acceptable tolerance ranges, typically ±5 HU for water
• Uncalibrated scanners introduce systematic bias that propagates into all downstream radiomic features, particularly first-order statistics like mean and median intensity

HU rescaling directly affects the reproducibility of quantitative imaging biomarkers:

• First-order features (mean, median, standard deviation) are linearly sensitive to rescaling errors
• Texture matrices (GLCM, GLRLM) are affected when intensity discretization bin widths interact with rescaling precision
• Multi-center trials require ComBat harmonization or similar batch-effect correction when scanners use different rescaling implementations
• The Image Biomarker Standardisation Initiative (IBSI) mandates reporting of rescaling slope and intercept parameters to ensure computational reproducibility
• Features derived from absolute HU values (e.g., mean HU of a lesion) are more vulnerable to inter-scanner variability than relative texture features

The presence of iodinated contrast agents fundamentally alters HU values and complicates rescaling:

• Iodine has a high atomic number (Z=53), producing strong photoelectric absorption and elevated HU values
• Contrast-enhanced tissues can shift by +50 to +300 HU depending on vascularity and timing
• Multi-phase protocols (arterial, portal venous, delayed) yield different HU ranges for the same tissue
• Radiomic models must specify whether features were extracted from contrast-enhanced or non-contrast acquisitions
• Virtual non-contrast techniques attempt to computationally subtract contrast effects, but introduce their own rescaling artifacts

When a voxel straddles two tissue types with different attenuation coefficients, the resulting HU value represents a weighted average:

• Boundary voxels between bone and soft tissue produce intermediate HU values that don't correspond to any real tissue
• This effect is pronounced in thick-slice acquisitions (>3 mm) and low-resolution matrices
• Voxel resampling to isotropic dimensions partially mitigates partial volume effects before feature extraction
• Radiomic features from small lesions (<2 cm) are disproportionately affected due to higher surface-to-volume ratios
• Partial volume correction algorithms attempt to decompose mixed voxels but require assumptions about tissue composition

A statistical batch-effect correction method adapted from genomics to standardize radiomic feature values across different scanner vendors and acquisition protocols. ComBat is applied after feature extraction to remove residual technical variability that HU rescaling alone cannot address.

• Models scanner effects as additive and multiplicative noise terms.
• Preserves biological variability associated with the clinical endpoint.
• Often used in conjunction with rescaling for robust multi-center radiomic signature development.

Histogram-based metrics that describe the distribution of individual voxel intensities within a Volume of Interest (VOI). These features are directly dependent on accurate HU rescaling for cross-scanner comparability.

• Mean HU: Average radiodensity, sensitive to rescaling intercept errors.
• Entropy: Measures randomness of the intensity distribution.
• Skewness & Kurtosis: Quantify histogram asymmetry and peakedness, reflecting tissue heterogeneity.
• Unreliable if raw scanner units are not converted to standardized HU.

The broader computational framework for removing unwanted technical variability from radiomic features. HU rescaling is the foundational image-level harmonization step, while methods like ComBat operate at the feature level.

• Image-Level: Rescaling, voxel resampling, and intensity normalization.
• Feature-Level: ComBat, quantile normalization, and nested autoencoders.
• A multi-stage harmonization pipeline is critical for building generalizable radiomic signatures that validate across external cohorts.

The interpolation of medical image data to create isotropic voxels—cubes with equal dimensions in all three spatial axes. This preprocessing step is performed alongside HU rescaling to ensure spatial consistency.

• Converts anisotropic acquisition voxels (e.g., 0.7 x 0.7 x 3.0 mm) to a uniform grid (e.g., 1 x 1 x 1 mm).
• Uses B-spline or linear interpolation to estimate intensity values at new grid points.
• Prevents rotationally variant texture features and ensures Shape Features are physically meaningful.