Feature Harmonization

An empirical Bayes batch-effect correction framework adapted from genomics. It adjusts feature values by modeling additive and multiplicative site effects while preserving biological covariates like tumor grade or patient age.

• Location Adjustment: Shifts the mean of each feature per batch
• Scale Adjustment: Rescales variance to match a reference batch
• Covariate Preservation: Retains associations with biological variables of interest

An independent international collaboration providing consensus-based reference standards for radiomic feature computation. IBSI defines exact mathematical formulas, preprocessing steps, and reporting guidelines to eliminate algorithmic variability.

• Benchmark Datasets: Provides digital phantoms for software validation
• Feature Naming Conventions: Standardizes terminology across platforms
• Reporting Checklists: Ensures reproducible methods sections in publications

Methods that rescale pixel values to a common range before feature extraction, mitigating scanner-dependent intensity drift.

• Histogram Matching: Warps an image's intensity distribution to match a reference template
• Z-Score Normalization: Transforms intensities to zero mean and unit variance per volume of interest
• White Stripe Normalization: Uses normal-appearing tissue as an internal reference for MRI intensity standardization

Neural network-based approaches that learn scanner-invariant representations directly from raw images, bypassing handcrafted feature engineering.

• Adversarial Training: Uses a domain discriminator to force the encoder to produce scanner-agnostic features
• CycleGAN Harmonization: Translates images between scanner domains without paired training data
• Feature Disentanglement: Separates imaging content from acquisition style in latent space

A filtering strategy that retains only features demonstrating high stability across test-retest scans and inter-scanner variations. Unstable features are discarded before model training.

• Concordance Correlation Coefficient (CCC): Measures agreement between repeated measurements
• Dynamic Range Thresholding: Removes features with biologically implausible variance
• Cluster Analysis: Groups features by stability profile to identify robust clusters

Quantitative methods to verify that harmonization successfully removed batch effects without eliminating true biological signal.

• Principal Component Analysis (PCA): Visual inspection of batch clustering before and after correction
• Silhouette Score: Measures batch mixing quality post-harmonization
• Biological Signal Preservation: Confirms clinical endpoint associations remain significant after adjustment

The process of converting continuous image intensity values into a finite number of discrete bins, a critical pre-processing step for texture matrix calculation. Discretization parameters directly impact feature reproducibility. Common approaches include:

• Fixed bin number: Divides intensity range into N equal bins
• Fixed bin width: Uses consistent bin size (e.g., 25 HU for CT) IBSI recommends fixed bin width for cross-scanner comparability.

The normalization of CT pixel values to a standardized scale where water = 0 HU and air = -1000 HU. This rescaling enables direct tissue density comparisons across different scanner models. Without HU rescaling, identical tissues scanned on different machines produce divergent intensity values, corrupting all downstream first-order and texture features. It is the foundational harmonization step for CT-based radiomics.

The process of interpolating medical image data to create isotropic voxels—cubes with equal dimensions in all three axes. Most clinical scans have anisotropic voxels (e.g., 0.5×0.5×3.0 mm), which distort shape and texture features. Resampling to a common resolution (typically 1×1×1 mm or 2×2×2 mm) ensures spatial measurements are consistent across acquisition protocols and scanner geometries.

A dimensionality reduction strategy that identifies and retains only radiomic features demonstrating high stability against test-retest and inter-observer variability. Features with intraclass correlation coefficient (ICC) below 0.75 are discarded. This filtering step eliminates non-reproducible features before harmonization, ensuring that only biologically meaningful—not scanner-dependent—signals enter predictive models.