Room: Track 2
Excellent repeatability (within same scanners) and reproducibility (across different scanners) of MRI are necessary for robust radiomics studies in radiotherapy. Using repeated brain MRIs of healthy volunteers across different MRI scanners, this study assessed repeatability/reproducibility of MR fingerprinting (MRF), compared to those of conventional contrast-weighted MRI.
Five volunteers were scanned three times on three separate days at each of two 3T scanners using 3D T1-weighted (T1W), 3D T2-weighted (T2W) and 3D MRF (resolution=1.2x1.2x3mm3, scan time=4.5/4/5.5 minutes respectively). Image registration and typical radiomics pre-processing steps, including bias-field correction and intensity normalization, were performed before calculating repeatability/reproducibility of 10 typical white/gray matter (WM/GM) structures. Effects of two most commonly used normalization methods, z-transformation with and without outlier removal, i.e. regulating voxels outside of the intensity range of mean±(3xstandard deviation) within the brain, on repeatability/reproducibility of T1W/T2W and MRF-T1/T2, were quantified.
Results from T1-contrast MRI are presented and similar results were observed for T2-constrast. Repeatability/reproducibility of raw MRF-T1 were 4.1±0.7%/5.7±1.0% for typical WM and 5.9±0.9%/8.0±2.0% for typical GM. Normalization did not affect these measures of MRF (Kruskal-Wallis p>0.3). In contrast, normalization significantly improved repeatability/reproducibility of T1W. Using z-transformation with additional outlier removal, normalized T1W had repeatability/reproducibility values of 2.4±0.5%/6.1±1.7% for WM and 4.4±2.1%/8.8±0.5% for GM, compared to 12.2±1.2%/52.5±7.0% for WM and 13.3±3.2%/55.2±4.7% for GM from the raw T1W. Detailed normalization steps implemented for pre-processing significantly affected achievable repeatability/reproducibility of normalized T1W/T2W. Without outlier removal, repeatability/reproducibility of intensity-normalized T1W can be significantly deteriorated (Wilcoxon signed-rank p<0.01), up to 4 folds in WM observed in this study (9.8±0.4%/20.5±1.8% for repeatability/reproducibility respectively).
Compared to T1W/T2W, 3D-MRF has more efficient acquisition and good repeatability/reproducibility that are independent of normalization and therefore, free of potential unwanted bias associated with normalization processes. Thus, MRF should facilitate more robust longitudinal or multi-center brain radiomics studies.
Funding Support, Disclosures, and Conflict of Interest: Yong Chen and Weili Lin is receive research support from Siemens Healthineers. Yong Chen. receives royalties from Siemens Healthineers.