Magnetic Resonance Spectroscopy Spectral Registration Using Deep Learning

Abstract

A novel convolutional neural network based spectral registration (CNN-SR) approach is introduced to achieve efficient and accurate simultaneous frequency-and-phase correction (FPC) of single-voxel MEGA-PRESS magnetic resonance spectroscopy (MRS) data. For this approach, one neural network was trained and validated using a published simulated and in vivo MEGA-PRESS MRS dataset with a wide-range of artificial frequency (0-20 Hz) and phase (0-90{degrees}) offsets applied. The proposed CNN-SR approach was subsequently tested and compared to the sequential FPC deep learning approaches and demonstrated more effective and accurate performance. Furthermore, a large random Gaussian signal-to-noise ratio (SNR 20 and SNR 2.5) and line broadening (0-20 ms) was introduced to the original simulated dataset to investigate our model performance compared to the other deep learning models. The testing showed that CNN-SR was a more accurate quantification tool and resulted in a lower SNR when compared with the other deep learning methods, due to having smaller mean absolute errors in both frequency and phase offset predictions. For Off spectra, the CNN-SR model was capable of correcting frequency offsets with 0.014 {+/-} 0.010 Hz and phase offsets with 0.104 {+/-} 0.076{degrees} absolute errors on average for unseen simulated data with SNR 20 and correcting frequency offsets with 0.678 {+/-} 0.883 Hz and phase offsets with 2.367 {+/-} 2.616{degrees} absolute errors on average at very low SNR (2.5) and line broadening (0-20 ms) introduced. In addition, we tested the simulated dataset with additional SNR and line broadening on a more refined model (CNN-SR+) where the pre-trained CNN-SR was further optimized by minimizing the difference between individual spectra and a common template with unsupervised learning. The performance on Off spectra was improved to 0.058 {+/-} 0.050 Hz for correcting frequency offsets and to 0.416 {+/-} 0.317{degrees} for correcting phase offsets. We further evaluated the ability of our model to process the published Big GABA in vivo dataset and CNN-SR+ achieved the best performance. Moreover, additional frequency and phase offsets (i.e., small, moderate, large) were applied to the in vivo dataset, and CNN-SR+ also demonstrated better performance for FPC when compared to the other deep learning models. These results indicate the utility of using deep learning for spectral registration and demonstrate the application of unsupervised learning in further improving the model to achieve state-of-the-art performance.