Kurtosis

A leptokurtic histogram exhibits a sharp central peak and heavy tails. In medical imaging, this indicates a large concentration of voxels tightly clustered around the mean intensity, punctuated by extreme outlier values.

• Clinical relevance: Often seen in largely homogeneous tumors with isolated necrotic cores or calcifications.
• Texture implication: Signals a tissue region that is mostly uniform but contains sparse, high-contrast micro-structures.
• Example: A meningioma with uniform enhancement but scattered psammomatous calcifications.

A platykurtic histogram features a flattened central peak and thin tails. This distribution suggests that voxel intensities are spread broadly across the region without a dominant central value or extreme outliers.

• Clinical relevance: Frequently observed in highly heterogeneous, infiltrative lesions where tissue density varies continuously.
• Texture implication: Indicates a lack of a single dominant tissue class; the region is structurally complex throughout.
• Example: A glioblastoma with diffuse infiltration and heterogeneous enhancement patterns.

Radiomic platforms typically report excess kurtosis, which is calculated as the raw kurtosis value minus 3. This normalization sets a standard Gaussian distribution to zero.

• Positive excess kurtosis: The histogram is more peaked than a normal distribution (leptokurtic).
• Negative excess kurtosis: The histogram is flatter than a normal distribution (platykurtic).
• IBSI compliance: The Image Biomarker Standardisation Initiative mandates reporting excess kurtosis to ensure cross-platform feature harmonization.

Kurtosis is highly sensitive to extreme outliers in the intensity distribution, making it a powerful but potentially unstable radiomic feature.

• Scanner noise impact: A single pixel with an erroneous extreme value due to CT beam hardening can artificially inflate kurtosis.
• Robustness testing: Kurtosis is often flagged during test-retest reproducibility analysis; only stable kurtosis values should enter predictive models.
• Mitigation: Applying intensity discretization and outlier filtering during pre-processing stabilizes the metric.

Kurtosis-derived signatures have demonstrated independent prognostic power across multiple oncological imaging modalities.

• Non-small cell lung cancer (NSCLC): Higher CT kurtosis values are associated with improved overall survival, potentially reflecting a more organized tumor architecture.
• Glioblastoma: Lower kurtosis on T1 post-contrast MRI correlates with aggressive, infiltrative phenotypes and shorter progression-free survival.
• Breast cancer: Kurtosis of dynamic contrast-enhanced MRI (DCE-MRI) kinetic maps helps differentiate benign from malignant lesions.

While both metrics describe histogram shape, they capture distinct biological phenomena and should not be treated as redundant.

• Kurtosis: Measures the extremity of deviations and the concentration around the mean. Driven by outlier micro-environments.
• Entropy: Measures the overall randomness and uniformity of the entire distribution. Driven by general tissue disorganization.
• Combined use: A tumor can exhibit high entropy (general chaos) but low kurtosis (no dominant core), or vice versa. Including both in a radiomic signature often improves model discrimination.

High kurtosis (leptokurtic distribution) indicates a sharp central peak with heavy tails, reflecting a tumor with a dominant homogeneous tissue component interspersed with extreme outliers like necrosis or calcifications. This is a hallmark of aggressive, poorly differentiated tumors.

• Leptokurtic (K > 3): Suggests a dominant cell population with scattered aberrant regions
• Platykurtic (K < 3): Indicates a flat, multimodal histogram typical of mixed-tissue or highly infiltrative lesions
• Mesokurtic (K ≈ 3): Approximates a Gaussian distribution, often seen in benign or well-differentiated tissue

Changes in kurtosis over time serve as a delta-radiomic biomarker for early treatment response. A shift from leptokurtic toward mesokurtic distributions during chemotherapy often indicates tumor homogenization and favorable response before volumetric shrinkage occurs.

• Decreasing kurtosis: Suggests necrosis resolution and tissue normalization
• Increasing kurtosis: May indicate treatment-resistant clones emerging as outliers
• Temporal analysis: Kurtosis changes detectable within 2-3 weeks of therapy initiation, preceding RECIST-measurable response

Kurtosis differentiates benign from malignant lesions and stratifies tumor grades. Glioblastoma multiforme consistently exhibits higher kurtosis values than low-grade gliomas due to its heterogeneous architecture of necrosis, hemorrhage, and cellular proliferation.

• High-grade gliomas: K values typically 4.5-7.2 on T1 post-contrast sequences
• Low-grade gliomas: K values typically 2.8-3.5, closer to Gaussian
• Lung nodules: Malignant nodules show significantly higher kurtosis than benign granulomas on non-contrast CT

Kurtosis serves as an independent prognostic factor in multiple cancer types. High kurtosis on pretreatment CT correlates with poorer overall survival in non-small cell lung cancer (NSCLC), reflecting intratumoral heterogeneity that drives drug resistance.

• NSCLC: Leptokurtic tumors associated with 40% higher hazard ratio for progression
• Colorectal cancer: Kurtosis of liver metastases predicts response to anti-angiogenic therapy
• Breast cancer: High kurtosis on DCE-MRI linked to triple-negative subtype and worse recurrence-free survival

Kurtosis is most powerful when combined with second-order texture features like GLCM contrast and GLRLM run-length non-uniformity. While kurtosis captures global histogram shape, texture matrices quantify spatial relationships, together providing a comprehensive heterogeneity profile.

• Kurtosis + Entropy: Distinguishes chaotic heterogeneity (high entropy) from structured heterogeneity (high kurtosis)
• Kurtosis + GLCM Correlation: Identifies whether outliers are spatially clustered or randomly distributed
• Multiparametric models: Combining kurtosis with skewness and entropy improves AUC by 0.08-0.12 over single-feature models

Kurtosis is moderately sensitive to acquisition parameters, requiring careful harmonization in multi-center studies. Slice thickness, reconstruction kernel, and contrast timing all influence the histogram shape, making ComBat harmonization essential for reproducible radiomic signatures.

• Slice thickness: Thinner slices (< 2mm) produce higher kurtosis due to reduced partial volume averaging
• Reconstruction kernel: Sharp kernels amplify noise, artificially inflating kurtosis
• IBSI compliance: Standardized intensity discretization with fixed bin width (25 HU) improves reproducibility

Measures the asymmetry of the intensity histogram distribution around the mean value. A positive skew indicates a longer tail on the right (high-intensity outliers), while a negative skew shows a tail on the left (low-intensity outliers). Unlike kurtosis, which focuses on tail extremity, skewness captures directional imbalance. In tumor imaging, a negatively skewed histogram often suggests necrotic regions with predominantly low-intensity voxels.

Quantifies the randomness or disorder in the distribution of voxel intensity values within a region of interest. High entropy indicates a broad, flat histogram with many intensity levels (heterogeneous tissue), while low entropy suggests a narrow, peaked distribution (homogeneous tissue). Kurtosis and entropy often correlate—high kurtosis with a sharp peak may indicate lower entropy, but heavy tails can simultaneously increase entropy, making their combined interpretation diagnostically powerful.

The second-order moment measuring the spread of intensity values around the mean. While kurtosis describes the shape of the distribution's tails and peak, variance quantifies overall dispersion. A distribution can have high variance but low kurtosis (uniformly spread), or low variance but high kurtosis (tightly clustered with occasional extreme outliers). Together, they differentiate between diffuse heterogeneity and focal anomalies in tissue texture.

The first-order central tendency representing the average voxel intensity within the ROI. Kurtosis is independent of the mean—two regions can share identical mean intensity but exhibit radically different kurtosis values. A tumor with mean intensity matching healthy tissue but elevated kurtosis may harbor calcifications or dense cellular clusters invisible to simple thresholding, making kurtosis a critical complementary feature in radiomic signatures.

A dimensionality reduction strategy that identifies radiomic features demonstrating high stability against test-retest and inter-observer variability. Kurtosis is particularly sensitive to outlier voxels from image noise or segmentation errors, making it a candidate for robustness filtering. Features with intraclass correlation coefficient (ICC) below 0.75 are typically excluded. Retaining only robust kurtosis measurements ensures the radiomic signature generalizes across multi-center imaging protocols.