Intact metabolite spectrum mining by deep learning in proton magnetic resonance spectroscopy of the brain

Abstract

Purpose: To develop a robust method for brain metabolite quantification in proton magnetic resonance spectroscopy ( 1 H-MRS) using a convolutional neural network (CNN) that maps in vivo brain spectra that are typically degraded by low SNR, line broadening, and spectral baseline into noise-free, line-narrowed, baseline-removed intact metabolite spectra. Methods: A CNN was trained (n = 40 000) and tested (n = 5000) on simulated brain spectra with wide ranges of SNR (6.90-20.74) and linewidth (10-20 Hz). The CNN was further tested on in vivo spectra (n = 40) from five healthy volunteers with substantially different SNR, and the results were compared with those from the LCModel analysis. A Student t test was performed for the comparison. Results: Using the proposed method the mean-absolute-percent-errors (MAPEs) in the estimated metabolite concentrations were 12.49% ± 4.35% for aspartate, creatine (Cr), γ-aminobutyric acid (GABA), glucose, glutamine, glutamate, glutathione (GSH), myo-Inositol (mI), N-acetylaspartate, phosphocreatine (PCr), phosphorylethanolamine, and taurine over the whole simulated spectra in the test set. The metabolite concentrations estimated from in vivo spectra were close to the reported ranges for the proposed method and the LCModel analysis except mI, GSH, and especially Cr/PCr for the LCModel analysis, and phosphorylcholine to glycerophosphorylcholine ratio (PC/GPC) for both methods. The metabolite concentrations estimated across the in vivo spectra with different SNR were less variable with the proposed method (~10% or less) than with the LCModel analysis. Conclusion: The robust performance of the proposed method against low SNR may allow a subminute 1 H-MRS of human brain, which is an important technical development for clinical studies. K E Y W O R D S brain, convolutional neural network, deep learning, metabolite quantification, proton magnetic resonance spectroscopy

Reconstruction of spectra from truncated free induction decays by deep learning in proton magnetic resonance spectroscopy

“…For tFID 1024 (null truncation), the CNN returns the input spectrum as it is. Therefore, our CNN may be combined with another CNN unit that takes part in metabolite quantification, such that those spectra from truncated FIDs are recovered prior to entering the next CNN, whereas those fully acquired spectra simply pass through spec CNN spec . As previously discussed, screening of the quality of the input data is important, such as for minimizing the mentioned possibility of processing the data contaminated by unwanted signal.…”

“…The rat brain spectra were simulated using an in‐house script written in Python (v.3.6). Briefly, the relative concentration ranges of the metabolites for normal rat brain were determined according to the literature (Supporting Information Table ) and divided by the total number of spectra to be simulated (N = 50 000) for each metabolite.…”

“…The metabolite spectra and macromolecule baseline were combined with the relative ratio of 1:0.9 . The SNR and linewidth of the combined spectra were adjusted by adding random noise and line‐broadening (Gaussian filter), respectively.…”

“…Using the U‐net shown in Figure as a basis, the initial learning rate (1 × 10 −5 − 1 × 10 −0 ), momentum (0.8‐0.99) of the stochastic gradient descent with momentum algorithm, L 2 regularization parameter (1 × 10 −10 − 1 × 10 −1 ), and the number of initial filters (32‐128) were Bayesian‐optimized on the training and validation sets for the 3 CNNs individually (number of objective function evaluations = 30, mini‐batch size = 16, 50 epochs, and learning rate being scheduled to drop by a factor of 0.1 at the 41st epoch). For the training of the networks, a stochastic gradient descent with momentum algorithm was used.…”

“…For direct comparison among the 3 CNNs, the CNN‐predicted data were analyzed in the frequency domain for all CNNs. Both the normalized mean squared error (NMSE = / , where S GT and S pred are the GT and predicted spectra, respectively) and the Pearson’s correlation coefficient (r) were estimated from the spectra in real mode and used as the measures of performance of the CNNs.…”