Quantitative Evaluation of Radiomics Reproducibility In Deep Learning and Asir CT Reconstructions: A Phantom Study
Abstract
Purpose
To compare the reproducibility of radiomics features between deep learning (DL) and adaptive statistical iterative reconstruction (ASIR) CT images under systematic acquisition parameter variations using the GE Revolution Apex scanner.
Methods
Radiomics features were extracted from CT images of an advanced electron density phantom reconstructed using DL and ASIR across controlled variations in kVp, noise index (NI), and slice thickness (ST) using a standardized radiomics pipeline. Feature reproducibility was quantified using the coefficient of variation (CV). Features were classified as stable (CV 10%). Global reproducibility for all features was evaluated using median and mean CV over the variations of each acquisition parameter and reconstruction method.
Results
DL reconstruction demonstrated superior feature reproducibility compared to ASIR across all acquisition parameters. For kVp, the median CV decreased from 1.77% (ASIR) to 0.76% (DL), and the mean CV decreased from 3.19% to 2.32%. Under NI variation, median CV decreased from 3.06% to 1.59%, and mean CV decreased from 4.70% to 2.78%. Slice thickness remained the dominant source of variability for both techniques, with mean CVs of 30.42% (ASIR) and 23.51% (DL). DL increased the number of stable features from 76 to 84 under kVp variation and from 70 to 84 under NI variation. Both methods yielded 15 stable features under ST variation. DL also reduced unstable features, particularly for NI (5, vs 12 for ASIR), demonstrating improved radiomics reproducibility across acquisition settings.
Conclusion
Deep learning reconstruction significantly improves CT radiomics reproducibility relative to ASIR across clinically relevant variations in kVp, NI, and ST. Although slice thickness remains the primary driver of radiomics instability, DL consistently reduces feature variability and increases the number of robust features. These findings support DL-based reconstruction as a more reliable platform for quantitative CT radiomics and the development of reproducible imaging biomarkers.