Magnetic Resonance Spectroscopy Deep Learning Denoising Using Few in Vivo Data

Magnetic Resonance Spectroscopy (MRS) is a noninvasive tool to reveal metabolic information. One challenge of H-1-MRS is the low Signal-Noise Ratio (SNR). To improve the SNR, a typical approach is to perform a large number of Signal Averaging (SA). The data acquisition time, however, is proportional to the number of SA accordingly. A complete clinical MRS scan takes approximately 10 minutes in a common setting with the number of SA of 128. Recently, deep learning has been introduced to improve the SNR, but mostly the simulated data were used as the training set. This may hinder the MRS applications since some potential differences, such as acquisition system imperfections, physiological and psychologic conditions may exist between the simulated and in vivo data. In this paper, a new scheme that purely used repeated samples of realistic data was proposed. A deep learning model, Refusion Long Short-Term Memory (ReLSTM), was designed to learn the mapping from the time domain data with low SNR (24 SA) to the one with high SNR (128 SA). Experiments on the in vivo brain spectra of 7 healthy, 2 tumor and 1 cerebral infarction region show that using SA of 24, only 20% of a common setting, the spectra denoised by ReLSTM can provide the estimated concentrations of metabolites with the reliability comparable to those of the high-SNR spectra obtained commonly with 128 SA. Furthermore, compared with the state-of-the-art Low-Rank (LR) denoising method, the ReLSTM achieves lower relative errors and the Cramer-Rao lower bounds in quantifying some important biomarkers. In summary, ReLSTM can perform high-fidelity spectral denoising of the spectra with the fast acquisition (24 SA), which is valuable to MRS clinical studies.

Automated summary

Magnetic Resonance Spectroscopy (MRS) is a noninvasive tool used to reveal metabolic information. H-1-MRS faces the challenge of low Signal-Noise Ratio (SNR), which can be improved by performing a large number of Signal Averaging (SA). Deep learning has been introduced to improve SNR, but mostly using simulated data. This paper proposes a new scheme using realistic data samples for deep learning-based denoising, achieving high-fidelity spectral denoising with fast acquisition.