Repeatability, robustness, and reproducibility of texture features on 3 Tesla liver MRI

Texture features are proposed for classification and prognostication, with lacking information about variability. We assessed 3 T liver MRI feature variability. Five volunteers underwent standard 3 T MRI, and repeated with identical and altered parameters. Two readers placed regions of interest using 3DSlicer. Repeatability (between standard and repeat scan), robustness (between standard and parameter changed scan), and reproducibility (two reader variation) were computed using coefficient of variation (CV). 67%, 49%, and 61% of features had good-to-excellent (CV ≤ 10%) repeatability on ADC, T1, and T2, respectively, least frequently for first order (19–35%). 22%, 19%, and 21% of features had good-to-excellent robustness on ADC, T1, and T2, respectively. 52%, 35%, and 25% of feature measurements had good-to-excellent inter-reader reproducibility on ADC, T1, and T2, respectively, with highest good-to-excellent reproducibility for first order features on ADC/T1. We demonstrated large variations in texture features on 3 T liver MRI. Further study should evaluate methods to reduce variability. •49–67% of texture feature measurements had good-to-excellent repeatability.•Only 19–22% of texture features had good-to-excellent robustness to parameter change.•Only 25–52% of texture features had good-to-excellent interreader reproducibility.•First order features were least repeatable, but were most frequently reproducible.•Different parameter changes had varying effects on texture feature outputs.