MR spectroscopy frequency and phase correction using convolutional neural networks
Dávid Ma, Hortense Le, Yuming Ye, Andrew F. Laine, Jeffery A. Lieberman, Douglas L. Rothman, Scott A. Small, Jia Guo
Abstract
PURPOSE: To introduce a novel convolutional neural network (CNN)-based approach for frequency-and-phase correction (FPC) of MR spectroscopy (MRS) spectra to achieve fast and accurate FPC of single-voxel MEGA-PRESS MRS data. METHODS: Two neural networks (one for frequency and one for phase) were trained and validated using published simulated and in vivo MEGA-PRESS MRS dataset with wide-range artificial frequency and phase offsets applied. The CNN-based approach was subsequently tested and compared to the current deep learning solution: multilayer perceptrons (MLP). Furthermore, random noise was added to the original simulated dataset to further investigate the model performance at varied signal-to-noise ratio (SNR) levels (i.e., 10, 5, and 2.5). Additional frequency and phase offsets (i.e., small, moderate, large) were also applied to the in vivo dataset, and the CNN model was compared to the conventional approach SR and model-based SR implementation (mSR). RESULTS: The CNN model is more robust to noise compared to the MLP-based approach due to having smaller mean absolute errors in both frequency (0.01 ± 0.01 Hz at SNR = 10 and 0.01 ± 0.02 Hz at SNR = 2.5) and phase (0.12 ± 0.09° at SNR = 10 and -0.07 ± 0.44° at SNR = 2.5) offset prediction. Furthermore, better performance was demonstrated for FPC when compared to the MLP-based approach, and SR when applied to the in vivo dataset for both with and without additional offsets. CONCLUSION: A CNN-based approach provides a solution to the automated preprocessing of MRS data, and the experimental results demonstrate the quantitatively improved spectra quality compared to the state-of-the-art approach.