Radiomics-informed CT protocol optimization through feature-level trade-off analysis in a phantom study

Robust radiomic features are critical for reliable quantitative CT analysis; however, their reproducibility is limited by variations in acquisition protocols. This study aims to (1) identify radiomic features predictive of low-contrast detectability, (2) quantify their sensitivity to acquisition parameters, and (3) integrate these factors through a feature-level trade-off framework for CT protocol optimization. A phantom containing 30 inserts was scanned across multiple CT protocols, and 107 radiomic features were extracted. Feature robustness was assessed using CV and ICC, followed by ComBat harmonization, which approximately doubled the proportion of reproducible features. A random forest model achieved strong performance (AUC = 0.94, accuracy = 0.90), with first-order features showing the highest predictive power. Sensitivity analysis using Spearman's correlation revealed that texture features were more influenced by protocol variations (ρ≈ 0.15-0.17), while shape features were more associated with dose (ρ≈ 0.13-0.14). A consistent trade-off was observed between detectability and sensitivity (-0.07 to -0.17), indicating that maximizing detectability alone is insufficient for protocol design. The proposed framework enabled systematic protocol ranking, demonstrating that moderate acquisition settings provide a more balanced and robust performance. These findings establish a quantitative foundation for radiomics-informed CT protocol optimization, with potential for further validation in clinical settings.