Purpose: Intra- and inter-scanner variations associated with MRI for radiomics and machine-learning can significantly affect robustness of study results. Using phantom measurements, this study quantitatively assessed intra-/inter-scanner variability and contrast-to-noise-ratio (CNR) of the magnetic resonance fingerprinting (MRF) technique, compared to those with conventional contrast-weighted MRI.
Methods: An MRI phantom with four different concentrations (25%/35%/40%/50%) of polyvinylpyrrolidone (PVP) solution (that has T1 characteristics similar to brain) was scanned using both conventional MPRAGE T1-weighted and an optimized 3D-MRF sequence. For both sequences, the PVP phantom (at all four PVP concentrations) was scanned five times on each of two 3T scanners to assess for both intra- and inter-scanner variability. For each PVP concentration, the magnitude of the intra- and inter-scanner variability was compared for the original MPRAGE T1-weighted and MRF-T1 MRIs, and between their intensity-normalized counterparts (image intensity was normalized by the mean intensity of the PVP sample of 50% concentration to simulate one of the most commonly used normalization approaches in radiomics/machine-learning). CNRs between different PVP concentrations were also calculated for normalized images.
Results: The intra- and inter-scanner variability of the MRF-T1 map was 2.2%Â±0.8% and 2.5%Â±0.5%, respectively; vs. 14%Â±5.0% and 41%Â±2.2% for the MPRAGE T1-weighted MRI. Although intensity-normalization did improve MPRAGE T1-weighted MRI intra- (5.3%Â±3.8%) and inter-scanner variability (3.2%Â±1.6%), both variability levels were significantly higher (p=0.029), up to 5-fold, compared to the normalized MRF-T1. CNR values of normalized MRF-T1 were, on average, 2~3 times higher than with the normalized T1-weighted MRI. The results were similar at all PVP concentrations.
Conclusion: Compared to conventional contrast-weighted MRI, 3D-MRF demonstrates significantly lower intra-/inter-scanner variability and better CNR. Thus, imaging data from the 3D-MRF technique should be more quantitatively consistent across time/scanners and should facilitate more robust multi-center radiomics and model-based machine-learning.