4242

A Time Series Analysis Method to Achieve Improved Water Suppression in 1H MRS

Keshav Datta^1,2 and Daniel Mark Spielman^1,2

¹Radiology, Stanford University, Stanford, CA, United States, ²Electrical Engineering, Stanford University, Stanford, CA, United States

Synopsis

Adequate suppression of water signal is vital in detecting low concentration metabolites in in-vivo proton spectroscopy acquisitions. This problem, which is exacerbated in the presence of multiplicative noise induced by physiological motion, is not addressed by current water and lipid saturation-based approaches. Here we use a time series modeling approach and show that signal estimation techniques are extremely effective in suppressing the highly correlated physiological noise component to achieve over three orders of magnitude in-vivo water suppression.

Introduction

Several proton spectroscopic editing techniques rely on the differences of successive frames to eliminate the baseline water signal, in order to detect a low concentration metabolite. Despite the inclusion of saturation pulses to suppress water (e.g. CHESS¹), the presence of respiratory and cardiac motion noise can result in insufficiently cancelled water signals, severely limiting the reliable detection of the metabolites of interest. One application requiring robust water suppression stems from the recent interest in the use of polarization transfer techniques to enhance the ¹H lactate signal from a hyperpolarized [2-¹³C]Pyruvate experiment². The presence of a large water signal overlapping the methine lactate doublet at 4.6ppm poses a serious challenge in the practical use of this method. Here we model the temporal evolution of the spectrum as a time-series for each spectral point and use classical estimation techniques to predict baseline water signal. We compare different curve fitting techniques and show that an auto-regressive model best predicts the signal in the presence of correlated multiplicative noise, resulting in significantly improved water suppression.

Methods

A Free-Induction-Decay (FID) sequence preceded by three 1000Hz CHESS water saturation preparation pulses¹ was used to acquire dynamic data (TR= 1s) from either brain or kidney of a 450g male Wistar rat using a custom built ¹H transmit-receive birdcage coil in a 3T PET/MR scanner (Signa, General Electric) for a total acquisition time of 300s. Each of the 300 temporal frames was reconstructed by applying a 1024-point Fast Fourier transform (FFT) to obtain a 5000Hz proton spectrum centered at the water resonance. To test the effectiveness of the water signal estimation method, every third frame was set to zero. This 300x1024 data set was then modeled as a time series for each spectral point, and the missing frames were reconstructed using linear interpolation, cubic spline interpolation, and auto-regressive modeling along the temporal dimension. The estimated water signal was then subtracted from the original frame to suppress water.

To test the efficiency of water suppression in the context of Lactate detection from polarization transfer experiment (reverse-INEPT²), a 100mM [2-¹³C]Lactate phantom was strapped to the back of a male Wistar rat (522g) and proton spectrum from a 30mm slice covering the kidney (including the phantom) was acquired as described previously using a custom built ¹H/¹³C dual-tuned birdcage T/R coil. Two 90⁰, 500us pulses spaced 1/2J=3.5ms apart, and 90⁰ phase difference, were played on the carbon channel just before the proton readout pulse starting 15th frame, every 5th frame to transfer the polarization from the C₂-carbon to the coupled methine proton. The three time series analysis methods were used to estimate water signal in the INEPT frames, and the transferred in-phase lactate doublet was recovered by subtracting the measured spectrum from the estimated one.

Results and Discussion

Time evolution of the proton spectrum acquired from a rat head for different spectral indices demonstrates the underlying correlated modulation due to physiological motion (Fig. 1). An auto-correlation based signal estimation technique is more efficient in predicting this pattern compared to linear and cubic interpolation as seen by the lower residue of water signal (Fig. 2). Utility of this water suppression method is demonstrated in Fig. 3 where the original water (SNR ~ 180,000) is almost completely eliminated. Fig. 4 shows the result of application of the proposed water suppression technique in polarization transfer experiment for Lactate detection. Here, signal from water and fat is eliminated to detect the in-phase methine doublet transferred from the coupled carbon, demonstrating the utility of this method in suppressing broadband background metabolites.

Conclusion

We propose a water suppression technique based on time series analysis that provides over three orders of magnitude of water signal reduction. We show that with this method a low concentration metabolite (methine lactate doublet) can be recovered from under a massive water peak. In general, this method can be applied for reconstruction of metabolite signals from a series of differential acquisitions.

Acknowledgements

The Lucas Foundation, GE Healthcare, and NIH grants R01 CA17683603, R01 EB01901802, R01 MH110683, and P41 EB01589121.

References

1. Haase A, Frahm J, Hänicke W, Matthaei D. ¹H NMR chemical shift selective (CHESS) imaging. Phys Med Biol. 1985 Apr; 30(4):341-4.

2. von Morze C, Reed GD, Larson PE, et al. In vivo hyperpolarization transfer in a clinical MRI scanner. Magn Reson Med. 2018;80:480–487. https://doi.org/10.1002/mrm.27154

Figures

Figure 1: Proton spectrum acquired from a rat head using a slice selective free-induction-decay (FID) sequence (N=1024, BW=5000Hz, Flip Angle = 90⁰, TR=1s, number of frames=300) is plotted as a function of the spectral index. Evolution of selected points on the spectrum are shown as a function of time on the right to demonstrate the effect of respiratory and cardiac induced multiplicative noise on the spectrum. The proposed time series analysis method exploits this correlation in the temporal dimension for effective water suppression.

Figure 2: The residual water signal from rat head as estimated using time series analysis. Comparison of three different estimation techniques: linear (dash-dot), cubic spline (dashed) and an auto-regressive model (dotted) demonstrates the superior ability of the auto-regressive technique to estimate water signal arising from physiological noise.

Figure 3: Comparison of the unsuppressed water signal (dashed) in the rat head with the residual from the time series analysis method (circles), demonstrates that the proposed method achieves a high degree of water suppression (SNR of unsuppressed water = 180,000, SNR of residue ~ 200).

Figure 4: Application of the proposed water suppression technique in a polarization transfer experiment for lactate detection. Proton spectrum acquired from kidney of male Wistar rat strapped with a thermally polarized [2-¹³C]Lac phantom (100mM), demonstrates water and fat signals (dashed line) are effectively suppressed by the time series analysis method to reveal the transferred methine [2-¹³C] Lactate doublet otherwise buried under the water signal (solid line).

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)

4242