Author: Morgan A. Daly, David J. Goodenough, Andrew M. Hernandez, John M. Hoffman, Joshua Levy, Michael F. McNitt-Gray, Ali Uneri, Bino Varghese 👨🔬
Affiliation: University of California, George Washington University, David Geffen School of Medicine at UCLA, Johns Hopkins Univ, University of Southern California, The Phantom Laboratory 🌍
Purpose: Quantitative imaging is affected by CT acquisition and reconstruction conditions, limiting robustness in multi-site or -scanner studies. This work aimed to develop a dataset that will enable rigorous investigation into the relationship between CT image quality and quantitative imaging.
Methods: We identified a set of N=200 CT protocols spanning a range of conditions used clinically for lung imaging comprised of (i) five scanner models across two manufacturers, (ii) five CTDIvol levels, and (ii) eight wFBP kernels. For each condition, we imaged (1) an image quality phantom and (2) a texture phantom for measuring radiomic features. From (1), we extracted contrast-to-noise ratio (CNR), modulation transfer function (MTF), and noise power spectrum (NPS) data for each protocol. From (2), we extracted 89 radiomic texture features from four texture modules for each protocol: interspersed (A) 0.010” acetal shavings and (B) 0.030” acetal shavings, (C) 3 mm diameter polypropylene spheres, and (D) fiberglass insulation.
Results: The protocols included yielded a wide range in image quality illustrated by their NPS and MTF curves. Calculated image quality metrics such as MTFf50 (the frequency where MTF falls to 50%) ranged from 0.28 to 0.89mm-1 (μ=0.54mm-1), CNR ranged from 0.20 to 8.23 (μ=0.54), and NPSf_peak (the frequency of the NPS peak), ranging from 0.16 to 0.71mm-1 (μ=0.43mm-1). Initial investigations illustrate the complex relationships between radiomic features and image quality. Some feature-image quality metric pairs appeared to have no relationship (e.g., mean value vs all metrics), while other appeared linear (variance vs area under the NPS) or exponential (NGTDM_Coarseness vs MTFf50, or GLCM_JointEnergy vs CNR).
Conclusion: We have developed a first-of-its-kind phantom dataset combining radiomic data with quantitative image quality data across 200 different CT protocols acquired over five scanner models. This will enable rigorous investigation into the relationship between CT image quality and quantitative imaging.