Sungtak Hong1 and Jun Shen1
1National Institute of Mental Health, National Institutes of Health, Bethesda, MD, United States
1National Institute of Mental Health, National Institutes of Health, Bethesda, MD, United States
The present study demonstrated that the benefit
of larger training data sizes could be marginal after reaching a threshold
number of datasets in training a convolutional neural network to restore degraded in vivo 1H MRS spectra.
Figure 1. Schematic overview
illustrating the generation of dataset and the proposed network architecture featuring
consecutive three convolutional blocks. A pair of 1D convolutional layer and batch
normalization layer act as a fundamental component with four times repetitions
for completing each block. Network training was conducted with pairs of ground
truth spectra and progressively degraded spectra while minimizing the mean
squared error using Adam optimization algorithm. Learning rate was set to 10-4.
Figure 2. Numerically calculated 1H
MRS spectra at low SNR (left column) and high SNR (right column). CNN-predicted
spectra, difference spectra (ground truth – predicted), and NMSE illustrate the
impact from using different training sizes in CNN.