Purpose: This study aimed to design a radiomic phantom, to extract radiomic features of multi-parametric MRIs, and to evaluate feature variability from different MR scanners and scan sessions.
Methods: We printed a radiomic phantom which had three grid objects using a 3D printer; a 3×3×3cm3 mesh cubic with a 2mm thickness, and two 5×5×5cm3 mesh cubic with 1mm and 2mm thicknesses. The three objects were placed in a water container prior to MR scans on a 1.5T Siemens MR scanner and a Philips MR equipped in Unity MR-Linac. Three MR scans were performed with a one-week interval on both scanners. Using Pyradiomics and 3D slicer, features of T1, T2 and ADC image sets were extracted using open software. Then, the radiomic features were compared to assess (1)intra-scanner variability between sessions and (2)inter-scanner variability between MR scanners. Assessment software was developed using Python 3, which read the calculated radiomic parameters, and analyzed how they were consistent between scanners and scan sessions. We analyzed only 112 numeric parameters out of 120 Pyradiomics parameters (19 first-order-statistics, 26 2D/3D shape-based, 70 gray-level-matrix and 5 neighboring-gray-tone-difference).
Results: In Philips T1 intra-scanner variability study, 58, 30, 16, and 8 features were within 0-5%, 5-10%, 10-15%, and 15-20% variation, respectively. Siemens T1 intra-scanner variability showed that 80, 8, 5, and 9 features were within 0-5%, 5-10%, 10-15%, 15-20% variation, respectively. The inter-scanner analysis from this T1 data showed Siemens produced consistent scan quality. However, T2, and ADC analysis were slightly different from the T1 results. Complete analysis is in the supporting data.
Conclusion: We successfully evaluated radiomic features using our 3D printed phantom and developed the analysis software. Those parameters showed large variation can be eliminated or assigned with low weights for building an outcome prediction model. Our analysis software can be used for other radiomic software.