A Comparison of Denoising Methods in Dynamic MRS Using Pseudo-Synthetic Data

Purpose: MR spectroscopy of dynamic systems is limited by low signal to noise. Denoising along a series of acquired spec- tra exploits their temporal correlation to improve the quality of individual spectra, and reduce errors in fitting metabolite peaks. In this study we compare the performance of several denoising methods. Methods: Six different denoising methods were con- sidered: SIFT (Spectral Improvement by Fourier Threshold- ing), HSVD (Hankel Singular Value Decomposition), spline, wavelet, sliding window and sliding Gaussian. Pseudo-synthetic data was constructed to mimic 31Phosphorus spectra from exercising muscle. For each method the optimal tuning pa- rameters were determined for SNRs of 2, 5, 10 and 20 using a Monte Carlo approach. Denoised data from each method was then fitted using the AMARES algorithm and the results compared to the pseudo-synthetic ground truth. Results: All six methods produced improvements in both fitting accuracy and agreement with the ground truth, com- pared to unprocessed noisy data. The least effective meth- ods, SIFT and HSVD, achieved around 10-20% reduction in RMS error, while the most effective, Spline, reduced RMS er- ror by 70%. The improvement from denoising was typically greater for lower SNR data. Conclusions: Indirect time domain denoising of dynamic MR spectroscopy data can substantially improve subsequent metabolite fitting. Spline-based denoising was found to be the most flexible and effective technique.