Intensity Discretization

Divides the intensity range of each individual Volume of Interest (VOI) into a constant number of bins, regardless of the absolute intensity values.

• Mechanism: The bin width adapts to the VOI's specific min-max range.
• Advantage: Ensures consistent matrix size across patients, which is critical for GLCM and GLRLM calculations.
• Disadvantage: Highly sensitive to outliers; a single bright pixel can compress the dynamic range of the rest of the tissue.
• Use Case: Preferred when comparing relative tissue heterogeneity within a single scan series.

Uses a constant, pre-defined intensity interval to create bins across the entire dataset or image.

• Mechanism: The bin width is absolute (e.g., 25 Hounsfield Units for CT). The number of bins varies per VOI based on its intensity range.
• Advantage: Preserves the physical meaning of intensity differences, making it essential for cross-scanner comparisons and ComBat harmonization.
• Disadvantage: VOIs with narrow intensity ranges may produce sparse, non-informative matrices.
• Standardization: The IBSI guidelines strongly recommend FBW for multi-center studies to maintain biological interpretability.

An iterative, data-driven algorithm that minimizes the mean squared error between original and discretized intensities.

• Mechanism: Dynamically adjusts bin boundaries to concentrate bins in high-density regions of the intensity histogram and widen them in sparse tails.
• Advantage: Maximizes information retention for a given number of bins, outperforming uniform methods in preserving first-order statistics.
• Disadvantage: Computationally more expensive and can produce non-intuitive bin boundaries that complicate biological interpretation.
• Relevance: Useful when the primary goal is signal compression for deep radiomics rather than explicit texture matrix calculation.

Assigns bin edges such that each bin contains approximately the same number of voxels from the VOI's intensity histogram.

• Mechanism: Uses quantile-based thresholds rather than uniform intensity intervals.
• Advantage: Guarantees that every bin is well-populated, eliminating the problem of empty or sparse bins that destabilize entropy and homogeneity calculations.
• Disadvantage: Destroys the linear relationship between intensity values, making it impossible to compare absolute tissue density across patients.
• Trade-off: Maximizes local texture contrast at the expense of global intensity calibration.

Discretization parameters are the single largest source of non-biological variance in radiomic pipelines.

• GLCM Sensitivity: Too few bins (<16) collapse distinct tissue patterns into identical gray-levels, erasing subtle heterogeneity signals.
• GLRLM Sensitivity: Too many bins (>256) fragment continuous runs of similar intensity, artificially reducing run length metrics.
• Reproducibility Crisis: Studies show that changing the bin count from 32 to 64 can alter feature values by over 40%, exceeding the effect size of the biological phenomenon being measured.
• Best Practice: Always report discretization parameters in compliance with the IBSI reporting guidelines to enable independent validation.

The sequence of pre-processing steps critically alters discretization outcomes.

• Correct Order: Apply Hounsfield Unit rescaling → voxel resampling to isotropic dimensions → intensity discretization.
• Rationale: Interpolation during resampling creates new, non-integer intensity values. Discretizing after resampling ensures these interpolated values are correctly binned.
• Filter Interaction: When using Wavelet or LoG filters, discretization must be applied to the filtered image, not the original, as the intensity distribution has been fundamentally transformed.
• Pitfall: Discretizing before resampling introduces partial-volume artifacts that propagate through all downstream shape features and texture matrices.

A texture matrix that counts the number of consecutive pixels with the same gray-level value in a specific direction to capture structural coarseness. Discretization bin width directly influences run length counts; coarse binning merges distinct intensities, artificially elongating runs and altering Short Run Emphasis and Long Run Emphasis metrics.

• Quantifies structural anisotropy
• Computed across 13 directions in 3D
• Highly sensitive to intensity quantization

A texture matrix that quantifies the size of connected regions of identical gray-level values, independent of directional orientation. Discretization critically affects zone size distributions; finer binning fragments homogeneous regions into smaller zones, while coarser binning merges distinct tissue compartments, potentially obscuring intra-tumor heterogeneity.

• Rotationally invariant by design
• Captures tumor granularity and necrosis patterns
• Used extensively in oncology radiomics

The computational process of removing unwanted technical variability from radiomic features caused by differences in scanner models, acquisition protocols, or reconstruction kernels. Discretization parameters interact with harmonization techniques like ComBat; inconsistent binning across sites can introduce batch effects that harmonization algorithms cannot fully correct.

• Requires phantom-based calibration
• ComBat adapts genomic batch correction to imaging
• Discretization must be standardized pre-harmonization