Author: Alex Lindgren-Ruby, Jeffrey M. Moirano, Joseph Everett Wishart 👨🔬
Affiliation: University of Washington 🌍
Purpose:
To evaluate polynomial fitting of signal intensity for signal-to-noise ratio (SNR) estimation from weekly QC ACR MRI phantom images and compare against other SNR estimation methods as well as low contrast detectability (LCD).
Methods:
The proposed SNR estimation method was validated/optimized using Monte-Carlo simulation, images of the ACR MRI phantom, a measured SNR reference standard, and comparison with LCD scores. The optimization was run on weekly MRI QC data from 15 scanners (3 vendors) across 1.5 and 3 T systems. Optimal region-of-interest radius and polynomial order were determined from SNR estimates, Bayesian Information Criterion, and deviations from normality of the fit residuals, and compared to SNR estimation from background. Reference SNR from a single system using repeated measures was compared to the proposed and other methods. A subset of MRI weekly QC LCD images were scored blindly by four raters. The correlation between the estimated SNR versus mean LCD score and interobserver agreement were measured using Spearman rank correlation and the standard deviation of LCD scores between observers, respectively.
Results:
Monte-Carlo results showed the RMSE SNR for 2nd order fitting of a 4th order polynomial across the signal area was <1%. The global optimal ROI radius was determined to be approximately 7 voxels using a 2nd order polynomial. For 3 T systems the departure from normality of the residuals markedly increased for larger radii at lower order. SNR using polynomial fitting did not correlate with SNR using background methods. The background method was highly biased relative to the proposed method. The correlation between SNR using polynomial fitting and LCD score was 0.752 (P < 0.0001). The mean LCD spoke standard deviation was 2.8 between observers.
Conclusion:
Using polynomial fitting to estimate SNR from the ACR MRI phantom may be a viable alternative to other SNR or CNR estimation methods.