Synopsis

Keywords: Data Processing, Spectroscopy

The use of deep learning models for frequency and phase correction of spectral data was evaluated to reduce processing time of 3D MRSI datasets acquired from patients with glioma. Models for frequency and phase offset estimation were trained for data prior to coil-combination. Separate models for baseline removal of coil-combined data were similarly trained and evaluated. The voxel-wise peaks for spectra processed using the standard approach were evaluated and compared to the spectra processed using the proposed deep learning approach. Compared to standard processing, deep learning processing produced spectra with comparable SNR and linewidths in a shorter processing time.

Introduction

¹H MR spectroscopic imaging (MRSI) is a powerful noninvasive technique that allows for the measurements of brain metabolites, such as N-acetyl-aspartate (NAA), choline-containing compound (Cho), and creatine (Cr). Lactate-editing allows the separation of lactate peaks¹, a biomarker for anaerobic metabolism, from lipid signals, which indicate necrosis or aliasing from subcutaneous tissue. Although previous studies have demonstrated the importance of using 3D lactate-edited MRSI for evaluating patients with glioma², one of the challenges for its clinical translation is long and complicated data processing. Artifacts caused by subject motion, lipid suppression, and water peak removal may result in frequency and phase offsets in the resulting spectra. While the use of phased-array receive coils can improve SNR³, the frequency/phase variations among channels may cause broadened metabolite peaks, reduced SNR, and distorted baseline after coil combination. Standard MRSI processing procedures include voxel-wise frequency and phase corrections, coil combination, and baseline removal prior to quantification^4,5,6. Recently, a deep learning (DL) approach for spectral frequency and phase corrections was introduced for offset estimation of individual transients before averaging⁷, greatly reducing processing time. In this work, we evaluated the use of DL models for voxel-wise frequency/phase correction and baseline removal for 3D lactate-edited MRSI datasets acquired from patients with glioma.

Methods

MRI and MRSI Data: A total of 37 datasets (35 training/validation, 2 test) were acquired from patients with glioma using an 8-channel phased-array head coil on 3T MR scanners (GE Healthcare, Waukesha, WI). A proton density-weighted gradient echo images for obtaining estimates of coil sensitivities and 3D lactate-edited MRSI datasets (TE=144ms, TR=1300ms, 1cm isotropic spatial resolution, matrix size = 16-18 x 16-18 x 16, flyback in the SI direction, 2 cycles (Edit-On/Edit-Off), total acquisition time ~10 min) were added to the standard brain tumor protocol.

Processing of spectral data and generation of DL training data: The spectral data were preprocessed with reordering to the rectilinear grid, apodization using a 4Hz Lorentzian filter, zero-filling to 1024 points, and Fourier transformation to spectral domain¹. Training data for DL models were created using standard processing (ST)^5,6, that utilized residual water peaks removed from Edit-OFF cycle spectra by HSVD filter⁸ for frequency/phase estimations. The phase/frequency-corrected spectra were then combined using an in-house software weighting with coil sensitivities⁹. ST included an additional round of frequency/phase estimation on the combined spectra³, not performed in DL processing. The ST estimated combined-spectra baseline was used as the training output for the DL model. The heights of Cho, Cr, and NAA were then estimated from baseline removed spectra with prior knowledge of peak locations.

DL network architecture and processing: DL networks are shown in Figure 1 and the DL-based processing workflow is summarized in Figure 2. Magnitude spectra were used as input to the frequency estimation model and real-valued spectra were used for the phase estimation model⁷. Baseline estimation models (Figure 1b) were trained on coil-combined data. Voxels within the excitation regions were aggregated from 35 datasets and used to train the DL models (Figure 3). All models were trained on a custom workstation with AMD 3900X 12-core CPU with a Nvidia Titan X GPU with Python 3.8 and Tensorflow 2.2.

Spectral analysis: ST and DL processing was performed on the remaining 2 datasets not used in training. Root-mean-square errors (RMSE) comparing DL to ST estimated offsets were calculated. Metabolite (NAA, Cho, Cr) SNR and linewidths⁵ were estimated from both processing methods. Pearson’s correlation coefficients were calculated to assess voxel-wise correlations between the two methods.

Results

Training/prediction times for DL models are summarized in Figure 3. Frequency/phase prediction times for one 8 channel 3D dataset (18x18x16x8-channel = 41,472 spectra) were <3 minutes each. Baseline predictions for one 3D dataset were <1 minute (18x18x16 = 5,184 coil-combined spectra). DL models successfully estimated frequency and phase in the test datasets for individual channel data (RMSE=7.9 for frequency. RMSE for phase was not included as the measurements are in degrees). An example of final ST and DL processed spectra are displayed in Figure 4b, showing agreement between the two methods. A comparison of the spectral analysis of the ST and DL processed spectra are shown in Figure 5. The SNR and linewidths values for all metabolites were correlated (SNR: R>0.9 all metabolites, linewidth: R>0.6 all metabolites) indicating the data processed with DL agreed well with ST.

Discussion and conclusion

Post-processing of spectral data can be time-consuming and complex. Recently introduced DL-based methods for frequency/phase correction both reduce processing time and simplify overall workflow. In this work, the use of a DL model for frequency and phase correction applied prior to coil combination was evaluated on patient data. We also created a DL model for baseline estimation, which is an important step before quantifying long TE MRSI. Although the ST processing pipeline included an additional frequency/phase correction step following coil combination, DL metabolite peak SNR and linewidths were similar to ST. Further improvements to the DL processing may be seen after a second such correction prior to baseline removal (see highlighted voxels in Figure 4b). Future work will include an additional model to estimate and remove the water peak along with the evaluation of more complex models for frequency and phase correction.

Acknowledgements

This project was supported by NIH T32 CA151022, R01 CA262630, P01 CA118816, P41EB013598, P50 CA097257, Department of Defense W81XWH2010292, and the Glioblastoma Precision Medicine Program.