A Voxel-Spacing-Aware Extension of PyRadiomics for Anisotropic Texture Analysis

September 15, 2026 | Source: arXiv AI

Tags: PyRadiomics, medical imaging, radiomics, CT, MRI, texture analysis, open-source

Researchers extend PyRadiomics to correctly handle anisotropic CT/MRI voxel spacing for texture feature extraction—covering all five major texture families (GLCM, NGTDM, GLRLM, GLDM, GLSZM) without resampling, with a backward-compatible release that exactly reproduces standard PyRadiomics under isotropic conditions.

Details

Radiomic texture features extracted from CT and MRI scans are routinely affected by anisotropic voxel spacing—where identical voxel offsets in different physical directions represent different distances. Standard PyRadiomics ignores this, introducing a systematic source of error in texture measurements used for downstream predictive models.\n\nThis extension implements physically correct anisotropic texture computation across five families without generating interpolated gray levels (which could themselves distort texture characteristics). GLCM uses anisotropy-relative angular aggregation; NGTDM uses weighted neighborhood averaging; GLRLM, GLDM, and GLSZM operate on a finite-volume zero-order-hold representation derived from the native anisotropic grid.\n\nValidation on synthetic 3D phantoms confirms exact reproduction of standard PyRadiomics under isotropic conditions across 75 texture features, and numerical distinction from nearest-neighbor, linear, and B-spline resampling approaches under anisotropic conditions. The extension is backward-compatible, making it a drop-in upgrade for existing PyRadiomics users.