PyRadiomics

The process of converting continuous image intensity values into a finite number of discrete bins, a critical pre-processing step for texture matrix calculation. PyRadiomics supports two primary strategies:

• Fixed bin number: Divides the intensity range into a user-specified number of equal-width bins
• Fixed bin width: Uses a constant bin size in Hounsfield Units or original intensity units

Discretization directly impacts GLCM, GLRLM, and GLSZM feature values. Coarse discretization reduces noise but may obscure subtle textural patterns, while fine discretization preserves detail at the cost of increased computational complexity.

The process of interpolating medical image data to create isotropic voxels, ensuring spatial measurements are consistent across all three dimensions. PyRadiomics applies B-spline interpolation to resample images to a user-defined voxel spacing before feature extraction.

Key considerations:

• Anisotropic acquisition: Clinical CT scans often have slice thickness 3-5x larger than in-plane resolution
• Rotation invariance: Isotropic resampling ensures shape and texture features are not biased by acquisition orientation
• Interpolation method: B-spline preserves intensity gradients better than nearest-neighbor or linear methods

PyRadiomics extracts seven distinct feature classes from medical images, each capturing different aspects of tissue characteristics:

• First-Order Statistics (18 features): Histogram-based metrics like entropy, kurtosis, and skewness
• Shape Features (14 features): 3D geometric descriptors including sphericity, surface-to-volume ratio, and maximum 3D diameter
• GLCM (24 features): Second-order texture capturing spatial relationships like homogeneity and contrast
• GLRLM (16 features): Run-length metrics quantifying coarseness and structural directionality
• GLSZM (16 features): Zone-size metrics for regional homogeneity independent of orientation
• NGTDM (5 features): Neighborhood difference metrics for local intensity variation
• GLDM (14 features): Gray-level dependence metrics for connected voxel patterns

PyRadiomics applies image filters before feature extraction to capture multi-scale and frequency-domain information not visible in the original spatial domain:

• Wavelet Transform: Decomposes images into 8 frequency sub-bands (LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH) using high-pass and low-pass filtering along each axis
• Laplacian of Gaussian (LoG): Applies edge-detection at multiple sigma values (e.g., 1.0, 3.0, 5.0 mm) to highlight regions of rapid intensity change
• Square/Squareroot: Simple intensity transformations that alter the dynamic range
• Exponential/Logarithm: Non-linear intensity mappings for contrast adjustment

Filtering dramatically expands the feature space, often yielding 1,000+ features from a single ROI.

A dimensionality reduction strategy that identifies and retains only radiomic features demonstrating high stability against test-retest and inter-observer variability. PyRadiomics users typically apply this post-extraction using:

• Intraclass Correlation Coefficient (ICC): Measures agreement between repeated measurements; features with ICC > 0.75 are considered robust
• Concordance Correlation Coefficient (CCC): Evaluates reproducibility across scanners or observers
• Coefficient of Variation (CoV): Quantifies relative variability; features with CoV < 10% are retained

This step is essential for building generalizable radiomic signatures that are not artifacts of acquisition noise or segmentation variability.