Integer Quantization As a Hidden Source of Bias In MRI SNR and QA
Abstract
Purpose
Quality assurance (QA) testing of multichannel phased-array RF coils on modern MRI scanners frequently produces images with very low background noise. While prior work has addressed bias arising from magnitude reconstruction and Rician statistics, the impact of integer quantization during image storage has received little attention. This work investigates quantization-induced bias in low-noise MRI images and its effect on commonly used noise estimation methods.
Methods
Analytical expressions were derived to quantify quantization-induced bias in noise estimates obtained using the standard deviation, mean, root mean square and NEMA subtraction methods. Numerical integration of Rayleigh and chi probability density functions was performed across a wide range of noise levels and coil channel counts to derive correction factors. These factors were validated using Monte Carlo simulations and experimentally with phantom scans acquired on GE, Siemens, and Philips 1.5 T systems using 1-, 8-, 18-, and 21-channel RF coils under varying reconstruction scaling factors (Siemens) and slice thicknesses (GE and Philips).
Results
Integer quantization produced substantial bias in noise estimates at low noise levels, with uncorrected errors ranging from 5% to 60%. Application of simple analytical correction factors reduced errors to below 1% for all methods except NEMA subtraction, which remained highly variable due to signal instability. The mean-based noise estimator demonstrated the most robust performance across the widest range of noise conditions. A practical method was developed to differentiate between rounding and truncation. GE uses rounding in software versions 12.X and earlier and truncation in later versions. Siemens consistently uses truncation. Philips uses truncation and may additionally apply rescaling.
Conclusion
Integer quantization is a significant and underrecognized source of bias in MRI SNR measurements, particularly in high-SNR QA protocols. The proposed correction factors substantially improve the accuracy and reproducibility of noise estimates and can be readily integrated into routine MRI QA workflows.