Entropy

Entropy measures the uncertainty or randomness in the distribution of voxel intensity values within a Volume of Interest (VOI). A higher entropy value indicates a more heterogeneous and chaotic internal tissue architecture, often associated with aggressive tumor biology. It is derived solely from the first-order histogram, ignoring the spatial arrangement of pixels.

• Formula Basis: H = -Σ p(i) * log₂(p(i))
• p(i): Probability of occurrence of intensity level i
• Range: Typically normalized between 0 (all same intensity) and 1 (maximum randomness)

In oncological imaging, high entropy is a quantitative surrogate for intratumoral heterogeneity. This metric captures variations caused by necrosis, hemorrhage, cellular density changes, and angiogenesis that are often invisible to the naked eye. Studies have linked elevated entropy values in CT and MRI scans to poorer prognosis, treatment resistance, and higher tumor grading in lung, brain, and colorectal cancers.

• Necrosis: Creates chaotic low-intensity pockets
• Angiogenesis: Generates irregular contrast enhancement
• Cellularity: Dense packing alters water diffusion

The absolute value of entropy is highly sensitive to the bin width chosen during the intensity discretization pre-processing step. Using too few bins (e.g., 8) collapses the histogram and artificially lowers entropy, while too many bins (e.g., 256) introduces noise that inflates randomness. The Image Biomarker Standardisation Initiative (IBSI) recommends a fixed bin number (e.g., 32) or a fixed bin width (e.g., 25 HU) to ensure cross-study reproducibility.

• Fixed Bin Number: Standardizes feature scale across different dynamic ranges
• Fixed Bin Width: Preserves the physical meaning of intensity differences
• Reproducibility: Discretization is the primary source of entropy variance

While entropy is a first-order metric, it is often confused with second-order texture features. Entropy analyzes the frequency of intensity values, not their spatial relationships. In contrast, Gray-Level Co-occurrence Matrix (GLCM) Entropy (a distinct feature) measures the randomness of neighboring pixel pairs. A tumor can have high histogram entropy but structured spatial patterns, or vice versa.

• First-Order Entropy: Ignores pixel location; pure histogram analysis
• GLCM Entropy: Measures joint probability of adjacent pixel pairs
• Complementary: Both are often used together in radiomic signatures

Entropy is generally considered a moderately robust feature against test-retest variability compared to higher-order textures. However, it remains vulnerable to noise and reconstruction kernel changes in CT imaging. Smooth kernels reduce entropy by blurring intensity transitions, while sharp kernels increase it. Feature harmonization techniques like ComBat are often applied to mitigate these scanner-specific biases in multi-center trials.

• Kernel Effect: Sharp kernels inflate entropy values
• Slice Thickness: Thicker slices average partial volumes, reducing entropy
• Harmonization: ComBat removes unwanted technical variability

Entropy rarely acts as a standalone biomarker but is a critical component of multivariate radiomic signatures. It is often combined with GLCM Homogeneity (inverse correlation) and Skewness to build predictive models for lymph node metastasis or immunotherapy response. Dimensionality reduction techniques like Principal Component Analysis (PCA) frequently retain entropy due to its high variance and independence from volume.

• Signature Example: Entropy + Homogeneity + Sphericity for NSCLC survival
• Delta-Radiomics: Changes in entropy over time indicate treatment response
• Feature Selection: Entropy often survives robustness filtering

The family of histogram-based metrics to which entropy belongs. These features describe the distribution of individual voxel intensities within a region of interest without considering spatial relationships.

• Key members: Mean, median, standard deviation, skewness, kurtosis, and entropy
• Entropy's role: Quantifies the shape of the histogram—high entropy indicates a broad, flat distribution of intensities; low entropy indicates a narrow, peaked distribution
• Clinical relevance: Often used as a baseline measure of tissue heterogeneity before applying more complex spatial texture matrices

A first-order statistical measure of the 'peakedness' or tailedness of the intensity histogram. Kurtosis and entropy are complementary descriptors of histogram shape.

• High kurtosis: A sharp peak with heavy tails, indicating most voxels cluster tightly around the mean with occasional extreme outliers
• Low kurtosis: A flatter distribution with lighter tails
• Relationship to entropy: A high-kurtosis distribution typically exhibits lower entropy due to concentration around the mean, while low kurtosis often correlates with higher entropy
• Example: In glioblastoma imaging, high kurtosis in contrast-enhanced regions may indicate a more homogeneous tumor core

A first-order measure of the asymmetry of the intensity histogram around the mean value. Skewness and entropy together characterize the overall shape of voxel intensity distributions.

• Positive skew: The tail extends toward higher intensity values; the mass of the distribution is concentrated on the left
• Negative skew: The tail extends toward lower intensity values; the mass is concentrated on the right
• Interaction with entropy: A highly skewed distribution may still exhibit high entropy if the spread of values is broad, or low entropy if values are tightly clustered
• Diagnostic value: Skewness changes in tumor regions can indicate necrosis or cystic degeneration

The critical pre-processing step of converting continuous image intensity values into a finite number of discrete bins. Discretization directly impacts entropy calculation.

• Bin width effect: Too few bins artificially suppress entropy by merging distinct intensity values; too many bins inflate entropy by treating noise as meaningful variation
• Fixed bin number vs. fixed bin width: Two competing discretization strategies standardized by the Image Biomarker Standardisation Initiative (IBSI)
• Entropy sensitivity: Entropy is among the radiomic features most sensitive to discretization parameters, making harmonization essential for multi-center studies
• Best practice: IBSI recommends a fixed bin width of 25 Hounsfield Units for CT entropy calculations

A second-order texture matrix that quantifies how often pairs of pixels with specific values occur in a defined spatial relationship. GLCM extends beyond first-order entropy by incorporating spatial information.

• GLCM Entropy: A distinct metric from first-order entropy—it measures the disorder of co-occurrence patterns, not just intensity distributions
• Key difference: First-order entropy ignores spatial relationships; GLCM entropy captures textural randomness by analyzing neighbor-pair probabilities
• Clinical example: A tumor with heterogeneous intensity but regular, repeating texture patterns may show high first-order entropy but low GLCM entropy
• Directionality: GLCM entropy can be computed at multiple angles (0°, 45°, 90°, 135°) to detect anisotropic tissue structures

The computational process of removing unwanted technical variability from radiomic features caused by differences in scanner models, acquisition protocols, or reconstruction algorithms.

• Entropy vulnerability: First-order entropy is highly susceptible to scanner-induced variations in noise texture and dynamic range
• ComBat harmonization: A statistical batch-effect correction method adapted from genomics that aligns feature distributions across imaging centers
• Impact: Without harmonization, entropy values from different scanners may reflect equipment differences rather than biological tissue properties
• Validation: Test-retest reproducibility studies consistently identify entropy as requiring harmonization before multi-center model deployment