Physics-Informed Evaluation of Dual-Energy CT Radiomics: Reproducibility and Electron Density Sensitivity In a Phantom Study
Abstract
Purpose
To evaluate the reproducibility and electron density sensitivity of CT radiomics features across spectral image types and deep learning reconstruction levels using a tissue-equivalent phantom.
Methods
A commercial electron density phantom with four tissue-equivalent inserts (adipose, blood, brain, cortical bone) was scanned using twelve spectral CT reconstructions, including 40 and 50 keV monochromatic images and Fat/Water material maps, each reconstructed with three deep learning levels using GE Apex Revolution. Radiomics features (n = 93 per protocol) were extracted using standardized PyRadiomics settings across first order and texture feature classes. Feature reproducibility across all reconstructions was quantified using the coefficient of variation (CV). Sensitivity to electron density was assessed using Spearman rank correlation, pooling all protocols and inserts. Features were considered reproducible if CV < 5% between all four inserts. Sensitivity to electron density was assessed using a Spearman rank correlation between mean feature values per insert and their corresponding relative electron density. Features were classified as physically meaningful if they were both reproducible and strongly associated with electron density (|rhoe | ≥ 0.8).
Results
Nine features demonstrated reproducibility across all spectral reconstructions for each insert, including first-order minimum and skewness, GLCM-based features (ClusterShade, Idmn, Idn, Imc1), GLDM DependenceEntropy, GLRLM RunEntropy, and GLSZM ZoneEntropy. Among these, six features exhibited strong monotonic dependence on electron density (|ρe| ≥ 0.8), including ClusterShade, Idn, Imc1, DependenceEntropy, RunEntropy, and ZoneEntropy. These features consistently increased or decreased with electron density across adipose, brain, blood, and bone, reflecting sensitivity to underlying material composition.
Conclusion
This phantom-based study identifies a physics-informed subset of spectral CT radiomics features that are invariant to acquisition protocol while exhibiting strong and interpretable dependence on electron density. These features represent robust candidates for quantitative spectral CT biomarkers and demonstrate the importance of jointly evaluating reproducibility and physical sensitivity in radiomics validation.