Radiomic Feature Reproducibility: The Impact of Inter-Scanner and Inter-Modality Variations

Radiomic features robustness analysis is a critical issue before clinical decision making. In this study, the reproducibility and robustness of radiomic features in computed tomography (CT) and magnetic resonance (MR) images of glioblastoma cancer patients were analyzed regarding inter-scanner and inter-modality variations.
CT and MR Images of eighteen glioblastoma cancer patients were used to extract the radiomic features following image segmentation. Coefficient of variation (COV), intraclass correlation coefficient (ICC), and concordance correlation coefficient (CCC) analysis were done to select the most robust features in all paired combinations of CT and MR images include T1-T2, T1-FLAIR, T1-ADC, T1-CT, T2-FLAIR, T2-ADC, T2-CT, FLAIR-ADC, FLAIR-CT, and ADC-CT.
The features with COV ≤ 5% or ICC ≥ 90% or CCC ≥ 90%, considered as the most robust features, include the shape features, Minimum (belong to first-order Features), IMC1, IDN, IDMN (belong to GLCM), and Run Length Non-Uniformity (belongs to Gray Level Run Length Matrix).
In this study we presented a large image feature variation among different imaging modalities including CT and MRI. Our results identified several robust features that could be used for further clinical analysis.