Zhipeng Cao1,2,3, Zhongliang Zu1,3, Kristin P. O'Grady1,3, Jun Ma2,3, William A. Grissom1,2,3, Seth A. Smith1,2,3, and John C. Gore1,2,3
1Radiology, Vanderbilt University, Nashville, TN, United States, 2Biomedical Engineering, Vanderbilt University, Nashville, TN, United States, 3Vanderbilt University Institute of Imaging Science, Vanderbilt University, Nashville, TN, United States
Synopsis
A method to accelerate chemical exchange saturation transfer (CEST) imaging is presented by reconstructing highly undersampled z-spectral image series using low rank compressed sensing. Results show x10 acceleration with an eight channel receiver with human brain imaging at 7T.
Introduction
Chemical exchange saturation transfer (CEST) imaging is a molecular imaging method in which contrast is derived by saturating multiple off-resonant frequencies to amplify the effects of exchange between protons of different chemical shifts. Amide proton transfer (APT) imaging, a specific variant of CEST, is of particular interest clinically for its ability to characterize tumors. However, two challenges that limit CEST for routine clinical applications are the relatively long scan time required to collect data at multiple saturation frequencies (a z-spectrum) and the need to quantify and correct the effects of B0 field inhomogeneities. This study presents a compressed sensing (CS) based method that undersamples and accelerates the data acquisition of a z-spectrum image series which addresses these challenges.
Theory
A method of applying CS to CEST imaging is proposed based on the fact that z-spectral images are typically highly correlated with singular value decomposition (SVD) and thresholding (Fig 1). Based on this, a CS algorithm that constrains the low rank cost function [1] ($$$\min_u{\bigl\{|L|_*\bigr\}}, s.t. E*u=v$$$, $$$|L|_*$$$ the nuclear norm of the low rank approximation ($$$L$$$) of z-spectrum image series $$$u$$$, $$$E$$$ the MRI Fourier encoding operator, $$$v$$$ the undersampled k-space data) should be more effective in reconstructing highly-undersampled data for the z-spectral image series than the conventional L1+TV based approach [2,3]. The use of a multi-element receive array can further accelerate CEST z-spectral data collection.
Method
The proposed method has been evaluated to date by retrospectively undersampling and reconstructing two datasets comprising images of (i) a rat with glioblastoma tumor, and (ii) a healthy human subject. The rat dataset was collected using a Varian DirectDrive 9.4T MRI scanner with a 38 mm Litz RF coil. The protocol includes a CEST preparation sequence with 5 s CW pulse with 0.25 uT power and 81 different saturation offset frequencies between +/- 5 ppm, each followed by a single-shot spin-echo EPI readout, with 64 x 64 matrix size, 30 x 30 mm FOV. The human dataset was collected using a Philips Achieva 7T MRI scanner with a sequence designed to quantify glutamate concentrations in gray and white matters. The protocol includes a CEST preparation sequence with 10 gaussian pulses each with 60 ms duration, 4.25 uT peak amplitude, 90% duty cycle, and a turbo gradient echo readout with 220 x 220 mm FOV, 1.5 x 1.5 x 10 mm voxel size, 4.1 ms TR, 2.7 ms TE, 10 degree flip angle, TFE factor of 11, and acquires images with 43 asymmetric saturation offset frequencies between +/- 5 ppm. Scan time of the fully-sampled single slice human subject data is 22 minutes. The undersampling for both dataset was performed with a series of one dimensional variable density cartesian trajectories. The effect of using same and different trajectories to undersample positive and negative saturation frequency pairs were also evaluated.
Results
Fig 1 demonstrates that z-spectral images are highly correlated and highly compressible by singular value decomposition and thresholding most of the singular values. The rat tumor dataset shows z-spectral and MTRasym information from x4.6 times undersampled data cannot be accurately reconstructed with L1+TV based CS method, but can still be accurately reconstructed with the proposed low rank CS method (Fig 2 & 3). Although use of asymmetric sampling patterns for positive and negative saturation frequencies improves the overall reconstruction accuracy due to undersampling artifacts being more incoherent, the artifact cancellation from image subtraction by using symmetric sampling patterns is better for MTRasym analysis. The healthy human subject dataset shows similar advantages of using a low rank based CS method compared to a conventional L1+TV based method, by reconstructing more accurate z-spectral images with less aliasing artifacts (Fig 4), and by reconstructing higher fidelity voxel-wise z-spectrum and MTRasym data compared to fully-sampled at various locations (Fig 5). Compared to previously-published results, the proposed method shows drastically enhanced compressibility, with a typical imaging acceleration ratio of x4.6 with single channel and x9.7 with eight channel receivers, compared to previously-reported x1.3~2 with single channel and x2.6~4 with 32 channel receivers [3].
Discussion & Conclusion
Here, a method to accelerate CEST z-spectral data acquisitions based on low rank compressed sensing is presented. The proposed method exploits the low rank feature of the z-spectral image series. Using symmetric undersampling patterns for positive and negative frequency pairs also improves the analysis of MTRasym. The presented results suggest CEST imaging can be highly accelerated and that whole brain coverage may be performed within clinically feasible scan time (<2 mins per slice). Ongoing work involves implementing the method with radial acquisition trajectories on a Philips 7 T human MRI scanner.
Acknowledgements
NIH R01 EB 016695 & U01 EB 025162References
[1] Otazo et al., MRM 2015. doi:10.1002/mrm.25240.
[2] Lustig et al., MRM 2007. doi:10.1002/mrm.21391.
[3] Heo et al., MRM 2017. doi:10.1002/mrm.26141.
