Systematic Approach to Identifying Sources of Variation in CT Radiomics: A Phantom Study

Variations in CT parameters affect extracted radiomic features, potentially confounding disease phenotype analysis in multicenter studies. However, the effects of such variations are not yet clearly understood. The purpose of this study was to assess CT radiomic feature robustness across various acquisition parameters using both a large water phantom and a liver phantom with 3D-printed textures. In this study, scans of both phantoms were acquired on Siemens CT with varied slice thicknesses, doses, voltages, and reconstruction methods. By varying these imaging parameters, 36 distinct images were acquired of the phantom. ROIs of the texture inserts were manually segmented and, from these, 107 features belonging to 6 subgroups of texture methods were extracted using Pyradiomics. A heat map was constructed to identify robust metrics. Across both phantoms, slice thickness showed the strongest impact on radiomic features (median <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf {\beta}=\mathbf{-0.18}$</tex>, IQR: -0.36 to 0.33; 0.6mm reference), followed by reconstruction algorithm (median <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{\beta}={-} {0.14}$</tex>, IQR: -0.36 to 0.31; FBP reference), and dose (median <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{\beta}={-} {0.10}$</tex>, IQR: -0.23 to 0.20; 40% reduction reference). Texture-based methods like GLCM, GLDM, GLRLM, and GLSZM demonstrated greater robustness compared to first-order measures such as SD, CV, and IQR. Our results may serve as a guide for clinicians and researchers alike to help them select radiomic metrics that are largely immune to these variations in image acquisition parameters and reconstruction methods. Radiomics-based phantom studies help identify stable radiomic features, essential for distinguishing genuine diagnoses from false positives caused by imaging parameter variations in clinical settings.