Magnetic Resonance Spectroscopy Spectral Registration Using Deep Learning

Abstract

Background: Deep learning-based methods have been successfully applied to MRI image registration. However, there is a lack of deep learning-based registration methods for magnetic resonance spectroscopy (MRS) spectral registration (SR).

Purpose: To investigate a convolutional neural network-based SR (CNN-SR) approach for simultaneous frequency-and-phase correction (FPC) of single-voxel Meshcher-Garwood point-resolved spectroscopy (MEGA-PRESS) MRS data.

Study type: Retrospective.

Subjects: Forty thousand simulated MEGA-PRESS datasets generated from FID Appliance (FID-A) were used and split into the following: 32,000/4000/4000 for training/validation/testing. A 101 MEGA-PRESS medial parietal lobe data retrieved from the Big GABA were used as the in vivo datasets.

Field strength/sequence: 3T, MEGA-PRESS.

Assessment: Evaluation of frequency and phase offsets mean absolute errors were performed for the simulation dataset. Evaluation of the choline interval variance was performed for the in vivo dataset. The magnitudes of the offsets introduced were -20 to 20 Hz and -90° to 90° and were uniformly distributed for the simulation dataset at different signal-to-noise ratio (SNR) levels. For the in vivo dataset, different additional magnitudes of offsets were introduced: small offsets (0-5 Hz; 0-20°), medium offsets (5-10 Hz; 20-45°), and large offsets (10-20 Hz; 45-90°).

Statistical tests: Two-tailed paired t-tests for model performances in the simulation and in vivo datasets were used and a P-value <0.05 was considered statistically significant.

Results: CNN-SR model was capable of correcting frequency offsets (0.014 ± 0.010 Hz at SNR 20 and 0.058 ± 0.050 Hz at SNR 2.5 with line broadening) and phase offsets (0.104 ± 0.076° at SNR 20 and 0.416 ± 0.317° at SNR 2.5 with line broadening). Using in vivo datasets, CNN-SR achieved the best performance without (0.000055 ± 0.000054) and with different magnitudes of additional frequency and phase offsets (i.e., 0.000062 ± 0.000068 at small, -0.000033 ± 0.000023 at medium, 0.000067 ± 0.000102 at large) applied.

Data conclusion: The proposed CNN-SR method is an efficient and accurate approach for simultaneous FPC of single-voxel MEGA-PRESS MRS data.

Evidence level: 4 TECHNICAL EFFICACY: Stage 2.

Keywords: CNN; MEGA-PRESS; MRS; deep learning; spectral registration; unsupervised learning.