Introduction

CEST imaging is currently limited by low SNR and long scan time.¹ Image reconstruction using deep learning (DL Recon) is capable of enhancing image signal-to-noise ratio (SNR) without losing image resolution or altering the image contrast. However, the influences of the deep learning based denoising methods were not well evaluated in quantitative imaging, especially CEST MRI. The objective of this study was to investigate if the enhanced SNR from DL Recon can help CEST imaging with higher spatial resolution without increasing scan time and if DL Recon enables CEST imaging with parallel imaging without its typical loss of SNR.

Methods

A spherical phantom (QalibreMD, Inc.) was used to hold five vials of gelatin (Verisol B, Gelita, Inc.) with different pH values of 6.0, 6.5, 7.0, 7.5 and 8.0, along with vials of distilled water and vegetable oil. The phantom was scanned on a 3T GE whole body scanner (Discovery MR750) using an 8-channel head coil at room temperature. The CEST images were acquired using a single-shot fast spin echo (SSFSE) sequence with fat saturation.² A saturation power of 2.0 μT and a saturation time of 2000 ms were used. 29 saturation frequencies from -7 ppm to 7 ppm in equal steps were acquired to obtain Z-spectra. A reference image was acquired without CEST saturation. WASSR was used for field inhomogeneity correction.³ Six sets of CEST images were acquired using a matrix size of 64, 128, 256, 384, 512 without parallel imaging (ASSET) and 512 with ASSET = 2. The TR was set to minimum to shorter scan time. It had a range from 2338.8 to 8607.7 ms. TE was set to minimum too, with a range from 27.0 to 35.8 ms. The FOV = 220 mm × 220 mm and slice thickness = 2 mm for all phantom scans. Three glioma patients were also enrolled and imaged using the same sequence as for the phantoms except with a matrix size of 256 and a slice thickness of 5 mm. All patient scans were approved by the institutional IRB and with written informed consents. MTR_asym at 3.5 ppm was measured for CEST analysis.

In addition to the standard recon, a vendor supplied DL Recon was used to reconstruct images using the same CEST raw data. The network was based on a deep convolutional residual encoder network pre-trained to reduce image noise without a loss in spatial resolution.

Results

Figure 1 compared low resolution CEST images with and without DL Recon. DL Recon was noted to be crispier and reduced the Gibbs ringing artifact (Fig. 1a vs. 1d). Because of the high SNR at the low resolution, CEST signals and local variations appeared to be similar with and without DL Recon (Fig. 1). With lower SNR at high resolution, DL Recon denoised the images (Fig. 2a vs. 2d) and generated a MTR_asym map that had lower noise (Fig. 2b,c vs. e,f). Due to its high SNR, the DL Recon Z-spectra and MTR_asym at a pixel-level were smoother and had a quality comparable to the ROI averaged Z-spectra and MTRasym using standard recon (Fig. 2g, h). The MTR_asym maps relative to the matrix size were compared using standard recon and DL recon (Fig. 3a vs. 3b). As expected, the MTR_asym maps from standard recon became nosier as matrix size increased and with ASSET (Fig. 3a). DL Recon compensated the SNR loss, and the corresponding high resolution MTR_asym maps from DL Recon images were smoother (Fig. 3b). While the standard deviation increased with spatial resolution for standard recon, it remained relatively unchanged for all resolutions for DL recon (Fig. 4). The MTR_asym was noted to increase with matrix size (Fig. 4), possibly due to an increased TR that allowed more magnetization recovery from saturation. Due to the same effect, the MTR_asym decreased when ASSET was used because TR was halved with ASSET (Fig. 4 512 vs. 512 with ASSET). For the in vivo images, the MTR_asym map was noted to be substantially less noisy with DL Recon (Fig. 5).

Discussion and conclusion

Our study demonstrated that DL Recon was applicable to CEST imaging which is currently limited by SNR and long scan time. With the SNR gain, DL Recon enabled higher resolution CEST imaging without the increase in scan time, which can potentially improve voxel-wise analyses or assessments of small ROIs. DL Recon also enabled CEST imaging with parallel imaging without its typical SNR loss, which can largely shorten the scan time. In this study, the images were acquired using FSE sequences, which had higher SNR compared with gradient echo sequences. Thus, the noise could be more of a problem in gradient echo based sequences, in which case DL Recon can be very beneficial. These benefits are expected to improve the CEST quantitation and its performance in clinical applications.

Acknowledgements

This work was supported in part by the Odyssey Program and Cockrell Foundation Award for Scientific Achievement at The University of Texas MD Anderson Cancer Center (S.Z.).