SU‐G‐IeP1‐02: Automated SNR Analysis of MR Images Using Model‐Based Noise Determination

Abstract

Purpose:To develop a robust automated method of signal‐to‐noise (SNR) analysis from MR images using a single image acquisition approach and model‐based noise determination.Methods:Three consecutive MRI phantom scans were performed according to the SNR test procedures in the 2015 ACR MRI Quality Control Manual. This was repeated for several different RF coils and on MRI scanners from two different vendors. The three images from each RF coil were analyzed by the single‐image (SNR‐ACR1) and two‐image (SNR‐ACR2) SNR analysis methods also described in the ACR QC Manual. In addition the images were analyzed by the automated method (SNR‐AUTO). The SNR‐AUTO method utilizes an ImageJ macro file that was developed to 1) segment the phantom from the background region by means of a simple histogram‐based thresholding algorithm, 2) measure the signal from a ROI encompassing 80% of the phantom area, and 3) determine noise by fitting the histogram of the background region (devoid of the phase encode direction) to a noise distribution model established in the literature.Results:The accuracy of the SNR‐AUTO analysis method is demonstrated by it yielding SNR results that on average were −0.8% ± 6.7% of the SNR‐ACR1 method and 7.5% ± 8.6% of the SNR‐ACR2 method. The SNR‐AUTO method demonstrated a higher level of reproducibility (average COV = 0.32%) than that of the other two methods (average COV = 1.01% and 1.69% for SNR‐ACR1 and SNRACR2 respectively) both of which require manual placement of ROIs. An additional finding is that the model‐based noise determination is relatively insensitive to the presence of minor artifacts in the background region.Conclusion:The SNR‐AUTO method proved to be an accurate, highly reproducible, time‐efficient method for analysis of SNR in MR images and is insensitive to minor artifacts in the noise‐measuring region.