Energy-Dependent Reproducibility of Radiomic Features across Virtual Monoenergetic Dual-Energy CT Images: A Phantom Study
Abstract
Purpose
Dual-energy CT (DECT) enables the generation of virtual monoenergetic images (VMI), which are increasingly used in quantitative imaging applications. This study aimed to quantify how virtual monoenergetic energy selection influences the reproducibility of radiomic features in DECT using a phantom-based repeat-scan design with heterogeneous materials.
Methods
A DECT system (Revolution CT, GE Healthcare, Waukesha, WI, USA) equipped with Gemstone Spectral Imaging (GSI) Xtream and fast kV-switching between 80 and 140 kVp was used. Reproducibility measurements were performed using a tissue-equivalent phantom containing six cylindrical inserts (CIRS Model 062M, CIRS Inc.) representing water (1.0 g/cm³), lung (0.507 g/cm³), adipose (0.960 g/cm³), breast (0.990 g/cm³), muscle (1.060 g/cm³), and liver (1.070 g/cm³). Cylindrical regions of interest (ROIs) measuring 2.5 cm in length and 1.8 cm in diameter were delineated for each insert. Radiomic feature extraction was performed using the LIFEx software. A total of 106 radiomic features were analyzed, and feature reproducibility across virtual monoenergetic energy levels was quantified using the intraclass correlation coefficient (ICC), categorized as Grade A (>0.80), Grade B (0.50–0.80), and Grade C (<0.50).
Results
The water insert demonstrated the highest overall reproducibility, with a mean ICC value of 0.82 (SD = 0.25; 95% CI, 0.77–0.86) across virtual monoenergetic energies from 40 to 80 keV. Among feature categories, intensity-based features (n = 23) showed the strongest agreement, with a mean ICC of 0.91 (SD = 0.38), and 74%–96% of features classified as Grade A across the evaluated energy levels. In contrast, the NGTDM features (n = 5) exhibited the lowest reproducibility, with a mean ICC of 0.34 (SD = 0.17). The NGTDM_Contrast feature demonstrated the poorest agreement, with an ICC of 0.07 (SD = 0.02).
Conclusion
Radiomic feature reproducibility in dual-energy CT varies across virtual monoenergetic energy levels, with substantial differences observed among feature categories.