Predicting Radiomic Feature Variation Due to CT Protocol Using Measured Image Quality Metrics
Abstract
Purpose
Radiomic feature reproducibility across CT protocols is limited by sensitivity to acquisition and reconstruction parameters. This work investigates whether standard physical image quality (IQ) metrics can quantitatively predict radiomic feature variation, enabling scanner-independent harmonization.
Methods
We acquired CT scans of a radiomics texture phantom and an IQ phantom (Corgi, The Phantom Laboratory) using N=160 protocols across four scanner models and two vendors. Protocols varied dose (CTDIvol: 3-17mGy) and reconstruction kernels (8 per scanner). From the IQ phantom, we calculated 32 scalar metrics from the measured 3D modulation transfer function (MTF) and 3D noise power spectrum (NPS). From the texture phantom, we extracted 91 radiomic features (PyRadiomics) from ROIs placed on four texture objects. Four representative features (Intensity Mean, Kurtosis, GLCM-Autocorrelation, GLCM-IDN) were selected based on mutual information feature importance for texture discrimination and low inter-feature correlation. We fit polynomial regression models using ElasticNet to predict radiomic feature values from IQ metrics for each feature-texture pair (16 models total). Hyperparameters were optimized via grid search. Performance was evaluated using 3-fold cross-validation, reporting R2 (mean±SD across textures) and mean absolute percentage error (MAPE).
Results
IQ metrics strongly predicted radiomic feature variance across the diverse protocols. The mean test R2 over the four texture objects was 0.93±0.03 for Intensity Mean (MAPE: 1.2%), 0.90±0.05 for Kurtosis (MAPE: 2.0%), 0.94±0.02 for GLCM-Autocorrelation (MAPE: 2.5%), and 0.86±0.02 for GLCM-IDN (MAPE: 0.4%), indicating that the variation in texture features was largely explained by fundamental differences in spatial resolution (MTF) and noise texture (NPS). Across the 16 models, the overall most important predictor terms by coefficient weighting were MTFaxialf10*MTFoblf90, kurtosis(MTFaxial)*MTFoblf90, and MTFaxialf10fpeak.
Conclusion
Physical IQ metrics accurately predict radiomic feature variation caused by CT protocol differences. This provides a potential pathway for a quantitative, scanner-independent framework for prospective harmonization in multi-center radiomics studies.