Quantitative Estimation of Signal-to-Noise Ratio In Accelerated Magnetic Resonance Imaging
Abstract
Purpose
To evaluate signal-to-noise ratio (SNR) estimates from polynomial fitting of the ACR MRI phantom signal intensity using accelerated imaging techniques and AI-assisted reconstruction and compare against accepted SNR measurement methods based on repeated acquisitions.
Methods
Multiple dynamic images of the ACR MRI Large Phantom were acquired on a 3 T MRI (Philips MR7700; Philips Healthcare, The Netherlands) using the ACR T1-weighted sequence, a turbo spin-echo (TSE) sequence, a SENSE parallel imaging sequence, three strengths of compressed SENSE parallel imaging, and four strengths of a vendor AI-assisted reconstruction sequence using a 32-channel head coil. Geometric maps of SNR were generated from uniform phantom images using accepted methods described in NEMA standards publication MS 1-2008 as well as using polynomial fitting of voxel intensities. Residuals from polynomial fitting were assessed for normality using an Anderson-Darling test (p = 0.05) and pixels where the null hypothesis was rejected were interpolated using surrounding data. Agreement between methods was assessed via root mean squared error (RMSE) between SNR map values using variable ROI size and polynomial order.
Results
Agreement between NEMA and polynomial fitting measurement methods varied, with ACR T1, SENSE, and AI-assisted reconstruction methods showing smallest RMSE. Compressed-SENSE and super resolution (SR) AI methods showed the worst agreement of the images analyzed.
Conclusion
Performance of SNR estimates generated using polynomial fitting varies compared with accepted methods depending on the acquisition and reconstruction methods.