Intensity Discretization

Partitions the intensity range of the Region of Interest (ROI) into a fixed number of bins, ensuring every scan has the same matrix size regardless of absolute intensity values.

• Mechanism: Divides the ROI's min-to-max intensity range into N equal-width bins.
• Key Advantage: Guarantees consistent matrix dimensions for Gray-Level Co-occurrence Matrix (GLCM) and Gray-Level Run Length Matrix (GLRLM) calculations.
• Critical Limitation: Loses direct physical meaning of intensity; a bin in one scan does not correspond to the same Hounsfield Unit range in another.
• IBSI Guideline: The Image Biomarker Standardisation Initiative (IBSI) recommends a default of 32 or 64 bins for FBN to balance texture granularity and statistical noise.

Uses a constant intensity interval per bin, preserving the absolute physical meaning of voxel values across different scans and patients.

• Mechanism: Bins are defined by a fixed width in original intensity units (e.g., 25 Hounsfield Units for CT).
• Key Advantage: Maintains absolute intensity calibration, making features directly comparable across different scanners and acquisition protocols.
• Trade-off: Produces a variable number of bins depending on the dynamic range of each ROI, leading to inconsistent matrix sizes.
• Application: Essential for ComBat Harmonization workflows where preserving physical intensity scales is required to correct for scanner-specific effects.

The discretization parameter directly controls the sparsity and statistical reliability of second-order texture matrices.

• Too Few Bins: Over-averages fine textures, collapsing distinct tissue patterns into the same gray level and reducing Entropy sensitivity.
• Too Many Bins: Creates a sparse Gray-Level Co-occurrence Matrix (GLCM) with many zero entries, making features like Cluster Prominence statistically unstable.
• Optimal Range: Empirical studies show 16–128 gray levels provide a stable plateau for most radiomic features before matrix sparsity degrades reproducibility.
• IBSI Benchmarking: The IBSI provides reference values for texture features at specific bin counts to validate software implementations.

Pre-discretization intensity normalization is required to handle outliers and standardize the input range before binning.

• Outlier Filtering: Voxels outside a defined percentile range (e.g., 0.5%–99.5%) are often clamped to prevent extreme values from distorting the bin distribution.
• Rescaling Methods: Common approaches include Z-Score Normalization and min-max scaling to a fixed range before applying FBN discretization.
• Absolute vs. Relative: FBW discretization operates on absolute intensities and typically requires no rescaling, preserving the native physical units of the modality.
• Reproducibility: Inconsistent outlier handling is a primary source of Intraclass Correlation Coefficient (ICC) degradation in multi-center studies.

The binning strategy interacts with the directional nature of texture matrices to influence feature rotational invariance.

• Gray-Level Run Length Matrix (GLRLM) and Gray-Level Size Zone Matrix (GLSZM) are computed at multiple angular offsets and then averaged.
• Discretization Granularity: Coarser binning reduces sensitivity to minor intensity fluctuations caused by anisotropic voxel interpolation during rotation.
• IBSI Compliance: The IBSI mandates specific interpolation and discretization sequences to ensure that rotationally averaged features are reproducible across software platforms.
• Practical Impact: Features like Short Run Emphasis (SRE) and Zone Percentage (ZP) show higher test-retest reliability when discretization is applied before rotational aggregation.