Phantom-based benchmarking of CT radiomics: The impact of reconstruction kernel and exposure on feature robustness, edge distortion, and machine learning generalizability

Purpose: Computed tomography (CT) radiomics is limited by technical instability because acquisition and reconstruction differences can alter feature values and reduce model transferability. This study evaluated CT radiomic robustness under controlled reconstruction-kernel and exposure variation and tested whether phantom-derived stable features retained downstream utility under clinical domain shift. Methods: Radiomic features were extracted from standardized material-specific regions of the CC-Radiomics-Phantom-2 dataset. Robustness was assessed using coefficient of variation (CV) as the primary metric and Susceptibility Index (SI) as supplementary confirmation. The All, Stable, and Stable no-shape feature sets were transferred unchanged to the non-small-cell lung cancer (NSCLC)-Radiomics Lung1 dataset and evaluated under baseline, kernel-shift, and vendor-shift settings using four classifiers. Results: Reconstruction kernel was the dominant instability source, with 104 of 107 features showing higher median CV under kernel variation than exposure variation. CV/SI screening retained 27 Stable features, enriched for shape descriptors, while Stable no-shape contained 13 features after removing all shape features. Blooming/edge-distortion analysis showed kernel-associated volume changes in the rigid acrylic layer, indicating that reconstruction effects extended to boundary-derived measurements. In Lung1, stability filtering produced model-dependent effects: the Stable subset most consistently showed higher AUC for Logistic Regression, whereas balanced-accuracy gains were less consistent. Stable no-shape generally performed below the full Stable subset. Conclusion: Phantom-derived stability filtering provides a reproducibility-oriented screening step for CT radiomics, but its downstream value requires classifier-specific and external-domain validation.