3178

Denoising dynamic CrCEST in skeletal muscle following exercise using low rank tensor approximations
Neil Wilson1, Mark A Elliott2, Dushyant Kumar2, and Ravinder Reddy2
1Siemens Medical Solutions USA Inc, Malvern, PA, United States, 2CMROI, University of Pennsylvania, Philadelphia, PA, United States

Synopsis

Dynamic CrCEST allows for high resolution mapping of the creatine kinase reaction recovery kinetics following exercise. However, voxel averaging over large ROIs is usually done to get reliable fits because of high variation due to limited SNR. Here, we show that improved denoising utilizing low rank tensor approximations that exploit the full dimensionality of the scans allows for reliable fits over smaller volumes or even single voxels, making muscle response heterogeneity measurable.

Introduction

The creatine kinase reaction plays a vital role in muscle energetics as the conversion of phosphocreatine (PCr) into free creatine (Cr) accompanies the production of ATP as an energy source. For decades, this reaction has been probed using 31-P MRS1. Recently, it has been shown that the CrCEST technique is an alternative that offers sensitivity benefits allowing for higher spatial resolution acquisitions2. Due to the need for adequate temporal resolution to characterize recovery, limited saturation offsets are acquired, and no averaging can be done. The resulting MTRasym maps have low SNR at 3T and require spatial averaging over large ROIs in order to achieve reliable fits to recovery curves.

Recently, it has been shown that low rank tensor approximations can offer improved denoising in CEST3 compared to low rank matrix techniques such as SVD4 and PCA5. Here, we apply that concept and exploit the additional time domain in dynamic CrCEST to show improved denoising at higher dimensionality despite acquiring a limited number of offsets and time points.

Higher SNR maps can lead to more reliable fits over smaller ROIs or even individual voxels and can allow exploration of muscle response heterogeneity within muscle groups that has not previously been achievable using 31-P spectroscopy based methods.

Methods

All subjects were scanned under an approved Institutional Review Board protocol of the University of Pennsylvania with written informed consent obtained from each volunteer. The CrCEST experiment was performed following in-magnet plantar flexion with a pneumatically-controlled foot pedal in a 3T scanner (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany). Single slice images were acquired axially through the center of the calf muscle with resolution 1.3x1.3x10 mm3. Six saturation offsets were acquired at ±1.5, ±1.8, and ±2.1 ppm with each image acquired every 5 s for an overall temporal resolution of 30 s. Sixteen time points were measured for the recovery curves. Low rank tensor approximations computed as a Tucker decomposition6 were taken on the raw k-space data in two ways: 1) the saturation offsets and time course were considered as a single temporal dimension for a data set of size (kx,ky,nc,temp) = (128, 128, 15, 6*16), where nc is the number of coils, and temp is the off-tp dimension; 2) the saturation offsets and time course were kept as separate dimensions resulting in a tensor of size (kx,ky,nc,off,tp). In the first method, rank reduction was applied keeping the first (90, 90, 15, 10) singular values in each dimension. In the second method, the first (90, 90, 15, 4, 3) singular values were kept. Recovery rates were obtained from a three parameter monoexponential fitting of each pixel's time course.

Results and Discussion

Figure 1 shows the CrCEST MTRasym maps from a single subject following exercise reconstructed without enhanced denoising (top row) and with the proposed rank reduction methods (second and third row). Both rank-reduced reconstructions show similar performance over the spatial maps with considerable improvement compared to the original. Figure 2 shows the reconstructed recovery curves from a single voxel in the lateral gastrocnemius that show the exercise response. The R2 goodness of fit is 0.65 for the conventional method, 0.92 for rank-reduced Method 1 that uses a single temporal dimension, and 0.98 for rank-reduced Method 2 that uses separate offset and time points. Fits over the entire lateral gastrocnemius were similar and had recovery time constants of 72, 78, and 76 s for the three methods. However, fits over the medial gastrocnemius had recovery time constants of 79, 97, 72 s, with that coming from Method 1 begin notably higher. The spatial distribution of R2 values had a mean of 0.46 for the conventional method, 0.81 for Method 1, and 0.90 for Method 2 and is shown in Figure 3. Lastly, Figure 4 shows the singular values along the temporal dimensions described in Method 1 (temp) and in Method 2 (separate off and tp).

While both rank-reduced approximations show good denoising relative to the original data, only the second method in which the saturation offsets and time points are kept separate utilizes the full dimensionality of the data. This difference was not obvious in the spatial maps showing recovery but is more pronounced in the time course over a single voxel and in the R2 goodness of fit. This is despite both methods having similar overall rank reduction. Recovery time constants from a medial gastrocnemius ROI showed potential bias in the results from Method 1, though this requires further investigation.

Performing the CrCEST exercise experiment at 7T would offer advantages both in terms of SNR and relative shift toward the slow exchange regime compared to 3T. However, because of the greater availability of 3T scanners and still favorable exchange rates, 3T optimization remains an important and viable goal.

Conclusion

The low rank tensor approximation of dynamic CrCEST images shows improved denoising that leads to more reliable recovery kinetic characterization. Despite the limited numbers of time points and offsets, rank reduction exploiting the full dimensionality by considering offsets and time points separate dimensions showed the best performance. Single pixel recovery curves can be reliably fit, allowing for study of muscle response heterogeneity.

Acknowledgements

No acknowledgement found.

References

[1] Dawson, M.J., Gadian, D.G. and Wilkie, D.R., 1978. Muscular fatigue investigated by phosphorus nuclear magnetic resonance. Nature, 274(5674), pp.861-866.

[2] Kogan, F., Haris, M., Singh, A., Cai, K., Debrosse, C., Nanga, R.P.R., Hariharan, H. and Reddy, R., 2014. Method for high‐resolution imaging of creatine in vivo using chemical exchange saturation transfer. Magnetic resonance in medicine, 71(1), pp.164-172.

[3] Chen, L., Cao, S., Koehler, R.C., van Zijl, P.C. and Xu, J., 2020. High‐sensitivity CEST mapping using a spatiotemporal correlation‐enhanced method. Magnetic resonance in medicine, 84(6), pp.3342-3350.

[4] Döpfert, J., Witte, C., Kunth, M. and Schröder, L., 2014. Sensitivity enhancement of (Hyper‐) CEST image series by exploiting redundancies in the spectral domain. Contrast media & molecular imaging, 9(1), pp.100-107.

[5] Breitling, J., Deshmane, A., Goerke, S., Korzowski, A., Herz, K., Ladd, M.E., Scheffler, K., Bachert, P. and Zaiss, M., 2019. Adaptive denoising for chemical exchange saturation transfer MR imaging. NMR in Biomedicine, 32(11), p.e4133.

[6] Tucker, L.R., 1966. Some mathematical notes on three-mode factor analysis. Psychometrika, 31(3), pp.279-311.

Figures

MTRasym maps without denoising (top row), with denoising method 1 (middle), and with denoising method 2 (bottom)

Recovery time course for a single voxel in the lateral gastrocnemius. Without denoising (left), denoising method 1 (center), denoising method 2 (right).

Spatial distribution of the voxelwise R2 goodness of fit for conventional processing (left, mean 0.46), method 1 (center, mean 0.81), and method 2 (center, mean 0.90). Color axis is scaled 0 to 1.

Singular values along the single temporal dimension of method 1 (top) and along the separate offset dimension (bottom left) and time points (bottom right) of method 2.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
3178