Test-Retest Reproducibility

The single most significant source of variability. Reproducibility is highly sensitive to deviations in the imaging protocol.

• Tube Current (mAs) and Voltage (kVp): Variations alter image noise texture, directly corrupting GLCM and GLRLM features.
• Slice Thickness: Thicker slices introduce partial volume averaging, destroying the fidelity of Shape Features and fine textures.
• Reconstruction Kernel: Sharp kernels (e.g., Bone) amplify noise, while smooth kernels blur edges. Switching kernels between scans can render First-Order Statistics incomparable.
• Pitch and Field of View: Alter the spatial resolution and voxel dimensions, requiring strict protocol lock-down for longitudinal studies.

Biological noise sources that cannot be fully eliminated but must be assessed for their impact on feature stability.

• Respiratory Motion: In thoracic and abdominal imaging, inconsistent breath-hold depths displace Volume of Interest (VOI) boundaries, introducing non-biological texture variance.
• Cardiac Pulsation: Transmits pulsatile motion to adjacent structures (e.g., lung bases, liver dome), degrading NGTDM features that rely on local neighborhood intensity differences.
• Peristalsis: Uncontrolled bowel motion changes the shape and density distribution of abdominal organs between scans, severely impacting GLSZM reproducibility.
• Mitigation: Features robust against these artifacts are classified as Robust Features and prioritized during model building.

Computational reproducibility requires deterministic, locked-down image preparation steps before feature extraction.

• Voxel Resampling: Interpolation to isotropic voxels (e.g., 1x1x1 mm³) must use identical algorithms (linear vs. B-spline) to avoid introducing systematic spatial biases.
• Intensity Discretization: The number of discrete bins (e.g., 32, 64) fundamentally alters texture matrix probabilities. A fixed bin width strategy is recommended by the Image Biomarker Standardisation Initiative (IBSI) to ensure cross-scanner stability.
• HU Rescaling: For CT, absolute Hounsfield Unit calibration is critical. Features derived from absolute intensity values are meaningless if the scanner's water attenuation baseline drifts.
• Normalization: Applying z-score normalization or histogram matching can stabilize features but may mask genuine biological changes if over-applied.

The 'jagged edge' problem. Manual or semi-automated segmentation introduces inter- and intra-observer variability that propagates into feature values.

• Boundary Uncertainty: Tumors with infiltrative margins (e.g., glioblastoma) exhibit poor segmentation agreement, directly destabilizing Shape Features like sphericity and surface area.
• Margin Inclusion: Including or excluding peritumoral tissue alters the texture context, switching the biological meaning of the GLCM correlation feature.
• Automated Segmentation: Deep learning-based auto-contouring (e.g., nnU-Net) significantly improves reproducibility over manual tracing, but the specific model version must be locked as a variable.
• Test-Retest Metric: The Dice Similarity Coefficient between two segmentations is a prerequisite metric; features from volumes with Dice < 0.8 are typically excluded from analysis.

Not all feature classes are created equal. Empirical test-retest studies consistently rank feature families by their inherent robustness to noise.

• High Stability: First-Order Statistics (mean, median) and basic Shape Features (volume, maximum 3D diameter) show excellent Concordance Correlation Coefficients (CCC > 0.9).
• Moderate Stability: GLCM features like homogeneity and contrast are generally stable if discretization is fixed.
• Low Stability: GLRLM and GLSZM features, particularly those capturing high-frequency heterogeneity (e.g., Short Run Emphasis), are highly sensitive to noise and reconstruction kernel changes.
• Selection Strategy: A Robust Feature Selection filter, retaining only features with CCC > 0.85 on a test-retest phantom or patient cohort, is mandatory before building a Radiomic Signature.

Long-term reproducibility requires monitoring the physical state of the imaging hardware, not just the software.

• Detector Aging: Gradual degradation of CT detector elements introduces ring artifacts and non-uniform noise profiles that corrupt Wavelet Transform sub-bands.
• Magnet Shimming (MRI): B0 field inhomogeneity drifts over time, distorting geometry and intensity uniformity, which invalidates Shape Features and Entropy measurements.
• Quality Assurance Phantoms: Routine scanning of standardized phantoms (e.g., ACR CT phantom) allows for Feature Harmonization by mapping feature values to a stable physical reference.
• ComBat Harmonization: A statistical technique originally from genomics, ComBat Harmonization can retrospectively correct for known scanner batch effects, preserving biological variance while removing technical drift.

A statistical batch-effect correction method originally developed for genomics and adapted to radiomics to standardize feature values across multiple imaging centers. ComBat uses an empirical Bayes framework to:

• Estimate and remove additive and multiplicative scanner effects
• Preserve biological variability associated with the outcome of interest
• Harmonize features without requiring phantom calibration data When test-retest variability is dominated by multi-center scanner differences rather than true biological change, ComBat harmonization is essential for restoring reproducibility.

The process of converting continuous image intensity values into a finite number of discrete bins, a critical pre-processing step for texture matrix calculation. Key parameters include:

• Fixed bin count: Divides the intensity range into a set number of bins (e.g., 32, 64, 128)
• Fixed bin width: Sets a constant bin size in Hounsfield Units or original units Discretization choices dramatically impact test-retest reproducibility. Fixed bin width methods generally yield more stable texture features across repeat scans than fixed bin count approaches.

The process of interpolating medical image data to create isotropic voxels, ensuring spatial measurements are consistent across all three dimensions. Common resampling targets include:

• 1×1×1 mm³ for high-resolution analysis
• 2×2×2 mm³ for computational efficiency
• In-plane resolution matching to preserve native axial detail Non-isotropic voxels cause rotationally variant texture features. Resampling to isotropic dimensions is a prerequisite for achieving directionally stable GLCM and GLRLM features across repeat imaging sessions.

A dimensionality reduction strategy that identifies and retains only radiomic features demonstrating high stability against test-retest and inter-observer variability. Common stability metrics include:

• Concordance Correlation Coefficient (CCC): Measures agreement between repeat measurements; features with CCC > 0.85 are typically retained
• Intraclass Correlation Coefficient (ICC): Quantifies reliability across multiple raters or timepoints
• Coefficient of Variation (CV): Expresses within-subject variability as a percentage of the mean This filtering step eliminates non-reproducible features before model training, preventing overfitting to measurement noise.