Synopsis

We apply a temporal denoising algorithm to hyperpolarized MRI. The SIFT method filters out temporal frequencies with low amplitudes, reducing noise while preserving dynamic information. We demonstrate this in a bioreactor setting, introducing hyperpolarized [1-13C] pyruvate to cells and human subjects and observing the conversion of pyruvate to lactate.

Introduction

Methods

Data were Fourier transformed and a range of chemical shift frequencies selected to identify a noise region (0-150 ppm). The data were Fourier transformed in the temporal dimension and threshold was set at 2xSTD of the temporal spectra in the noise region; values below the threshold were set to zero. The value at the zero frequency was kept to preserve the baseline. Since real peaks have broad support, isolated values were removed using erosion and dilation functions. The resulting denoised data were inverse Fourier transformed to generate denoised dynamic spectra (Fig1).

Dynamic metabolic data were generated by introducing HP [1-13C] pyruvate to breast cancer cells, as previously described (2). [1-13C]pyruvic acid was polarized in a SPINlab hyperpolarizer (GE). ¹³C NMR data were acquired on a 1T NMR spectrometer (Magritek) and human data were acquired on a 3T MRI (GE).

Results

Fig2 shows data with good SNR (the lactate signal is on the order of the pyruvate hydrate signal. The temporal spectrum is shown for the lactate peak (B), with a tall, broad peak at the low frequencies. There are a few noise peaks visible at the higher frequencies of the original data that are reduced at in the denoised data. The reduction in noise makes the lactate peaks more easily distinguishable from the surrounding noise. Fig3 shows data with much lower signal in the lactate peaks. In this case, there are relatively larger contributions from noise in the temporal spectrum that are removed by the SIFT algorithm. The dynamics of the lactate production are more clearly observed following denoising.

The dynamics from cells treated with a control and an AKT inhibitor are shown in Fig4. The difference in lactate production is more clearly distinguished in the denoised data. The SNR of lactate in both the control and treated group is between 0 and 5. In the denoised data the lactate SNR is 12-30 in the control group and 7-16 in the treatment group.

In a test/re-test experiment in a human prostate, the polarization varied between the two injections (35% and 22%) resulting in differerent SNR in the two datasets. Following denoising, the two data sets show comparable SNR, and lactate peaks that were at the noise level become more conspicuous. This demonstrates that this denoising method can be successfully used in vivo and with spatially resolved spectra.

Discussion

The heights of large peaks (e.g., pyruvate and the hydrate) do not change; the improvement in SNR is due to the reduction in the amplitude of the noise. For smaller peaks (e.g., lactate), the amplitude of an individual time point may change do the smoothing of the temporal dynamic, though the largest effect on improving peak detection Is the reduction of the noise at the base of the peak. There is no broadening of the peaks as there is with spectral apodization because the filtering is applied in the temporal, not spectral domain. The improvement in SNR allows us to detect peaks with lower SNR, such as the natural abundance [2-13C]pyruvate in the cell data and bicarbonate in the human prostate data.

There may be an artifact at the edges of the dynamics due to the periodic nature of the discrete Fourier transform. This is most significant when there is a sharp transition, such as for the large pyruvate signal in the first time point, which results in an artificially lower signal in the early time points. Beginning the acquisition prior to the introduction of the hyperpolarized substrate and acquiring until the signal is near zero so there is a smooth transition at the boundaries can minimize the decrease in signal. Alternatively, the early time points could be excluded from quantification.

Conclusion

Acknowledgements

References

1 Doyle M, Chapman, BL, Balschi JA, Pohost GM. SIFT, a Postprocessing Method That Increases the Signal-to-Noise ratio of Spectra Which Vary in Time. J Mag Reson B. 1994;103:128-133.

2 Keshari KR, Kurhanewicz J, Jeffries RE, Wilson DM, Dewar BJ, Van Criekinge M, Zierhut M, Vigneron DM, Macdonald JM. Hyperpolarized 13C spectroscopy and an NMR-compatible bioreactor system for the investigation of real-time cellular metabolism. Mag Reson Med. 2010;63(2):322-329.

Chen AP, Cunningham CH, Ozturk‐Isik E, Xu D, Hurd RE, Kelley DA, Pauly JM, Kurhanewicz J, Nelson SJ, Vigneron DB. High‐speed 3T MR spectroscopic imaging of prostate with flyback echo‐planar encoding, J. Magn. Reson. Imaging 2007; 25:1288–1292.