Homogeneity

Homogeneity is calculated by summing the squared joint probability of pixel pairs, weighted by the inverse of their intensity difference. The formula is: Σᵢ,ⱼ [p(i,j) / (1 + |i - j|)]. When pixel pairs have identical values (i = j), the denominator is 1, giving maximum weight. As intensity differences increase, the weight decreases hyperbolically. This inverse weighting structure makes the metric highly sensitive to local transitions rather than global contrast.

A homogeneity value approaching 1.0 indicates a GLCM where high-probability entries are concentrated along the main diagonal. This signifies that neighboring pixels frequently share identical or near-identical gray-level values, corresponding to visually smooth, uniform tissue regions. Conversely, a value near 0.0 indicates dispersed entries far from the diagonal, characteristic of heterogeneous, high-contrast textures with abrupt intensity transitions.

Homogeneity and Contrast are inversely correlated GLCM features that measure complementary texture properties. While Contrast weights joint probabilities by the squared intensity difference (i - j)² to highlight large variations, Homogeneity weights them by 1 / (1 + |i - j|) to suppress them. A tissue region with high Contrast will inherently exhibit low Homogeneity. Analyzing both metrics together provides a complete picture of local uniformity versus global variation.

In oncological imaging, low Homogeneity values within a tumor volume of interest (VOI) are frequently associated with intratumoral heterogeneity, a hallmark of aggressive malignancy. This textural chaos can reflect underlying necrosis, hemorrhage, or disorganized angiogenesis. Studies have correlated low Homogeneity on CT and MRI with:

• Poorer overall survival in non-small cell lung cancer
• Higher histological grade in gliomas
• Treatment resistance in colorectal liver metastases

Homogeneity values are highly dependent on intensity discretization—the number of gray-level bins used to construct the GLCM. A smaller bin count (e.g., 8-16 bins) compresses the dynamic range, forcing more pixel pairs onto the diagonal and artificially inflating Homogeneity. A larger bin count (e.g., 64-128 bins) preserves fine intensity differences, reducing Homogeneity. The Image Biomarker Standardisation Initiative (IBSI) recommends using a fixed bin number (e.g., 32) to ensure cross-study reproducibility.

Homogeneity is computed for specific pixel-pair spatial relationships defined by a distance (d) and angle (θ). A single GLCM may be calculated for 0°, 45°, 90°, or 135° directions. To achieve rotational invariance, standard radiomic pipelines like PyRadiomics compute Homogeneity for all four directions and report the average value. This ensures the texture measurement is independent of the original scan orientation relative to the tissue structure.

Homogeneity serves as a discriminative texture feature for distinguishing benign from malignant lesions. Malignant tumors typically exhibit lower homogeneity values due to heterogeneous cellular proliferation, necrosis, and chaotic angiogenesis.

• Breast cancer: Lower homogeneity on dynamic contrast-enhanced MRI correlates with higher histological grade (Nottingham Grade III vs. I)
• Gliomas: IDH-wildtype glioblastomas demonstrate significantly reduced homogeneity compared to IDH-mutant lower-grade gliomas on T2-FLAIR sequences
• Lung nodules: Homogeneity extracted from non-contrast CT helps differentiate adenocarcinomas from granulomas with AUC values exceeding 0.85 in validation cohorts

Pre-treatment homogeneity measurements function as predictive biomarkers for therapeutic outcomes. Tumors with higher baseline homogeneity often indicate treatment-resistant phenotypes due to uniform cellular density and poor drug penetration.

• Rectal cancer: High homogeneity on pre-chemoradiotherapy T2-weighted MRI predicts poor pathological complete response (pCR), guiding decisions toward surgical intensification
• Immunotherapy: Homogeneous texture in baseline CT scans of non-small cell lung cancer correlates with reduced response to checkpoint inhibitors, potentially reflecting immune-excluded tumor microenvironments
• Delta-radiomics: An increase in homogeneity during early treatment phases often signals favorable histopathological response, capturing tumor necrosis and fibrotic conversion

Homogeneity excels at quantifying fibrotic tissue remodeling across organ systems. Fibrotic processes produce architecturally uniform extracellular matrix deposition, resulting in elevated homogeneity values.

• Liver fibrosis: GLCM homogeneity from ultrasound elastography or CT texture analysis correlates strongly with METAVIR fibrosis stages (F0-F4), offering non-invasive alternatives to biopsy
• Idiopathic pulmonary fibrosis: Homogeneity extracted from high-resolution CT parenchymal analysis tracks disease progression and predicts forced vital capacity decline over 12-month intervals
• Cardiac fibrosis: Late gadolinium enhancement MRI homogeneity differentiates ischemic from non-ischemic cardiomyopathy patterns, informing device therapy candidacy

Homogeneity features serve as imaging surrogates for underlying genomic alterations, enabling non-invasive molecular phenotyping through radiogenomic associations.

• Breast cancer molecular subtypes: Luminal A tumors exhibit higher homogeneity on mammography and ultrasound compared to HER2-enriched and triple-negative subtypes, reflecting their indolent proliferative biology
• EGFR mutation status: Lung adenocarcinomas harboring EGFR sensitizing mutations demonstrate distinct homogeneity signatures on CT compared to KRAS-mutant tumors, potentially guiding biopsy site selection
• MGMT promoter methylation: Homogeneous texture patterns in glioblastoma multiforme MRI associate with MGMT methylation status, a critical determinant of temozolomide sensitivity

Homogeneity provides independent prognostic value beyond conventional staging systems, enabling refined risk stratification for personalized follow-up protocols.

• Head and neck squamous cell carcinoma: Low homogeneity on pre-treatment contrast-enhanced CT independently predicts worse overall survival (HR 2.1, 95% CI 1.4-3.2) after controlling for TNM stage and HPV status
• Pancreatic ductal adenocarcinoma: Homogeneity extracted from the tumor-stroma interface on preoperative CT stratifies patients into distinct recurrence-free survival groups, informing adjuvant chemotherapy intensity
• Cervical cancer: Homogeneity features from apparent diffusion coefficient maps predict lymph node metastasis with superior sensitivity compared to conventional size criteria (short-axis diameter >10mm)

Homogeneity achieves maximum clinical utility when integrated with complementary texture features into composite radiomic signatures. Isolated homogeneity measurements risk oversimplifying complex tissue architecture.

• Prostate cancer PI-RADS: Combining homogeneity with entropy and contrast from multi-parametric MRI (T2, DWI, DCE) improves clinically significant cancer detection (Gleason ≥3+4) compared to PI-RADS v2.1 assessment alone
• Rectal cancer complete response: A radiomic signature incorporating homogeneity, GLRLM run percentage, and shape compactness achieves 94% negative predictive value for identifying patients eligible for watch-and-wait protocols
• Feature harmonization: ComBat harmonization across scanner vendors is essential before deploying homogeneity-based signatures in multi-center trials to eliminate vendor-dependent variability