Room: Exhibit Hall | Forum 8
Purpose: To utilize deep learning to reduce acquisition time for the localized correlated spectroscopy (LCOSY) method.
Methods: GAMMA simulation was used to simulate 2D LCOSY spectra for the following major brain metabolites: N-acetyl aspartate, creatine, choline, glutamate, glutamine, myo-inositol, and others. The acquisition parameters were as follows: TE = 30ms, tâ‚‚ points = 2048, tâ‚? points = 100, direct spectral bandwidth (SBWâ‚‚) = 2000Hz, and indirect spectral bandwidth (SBWâ‚?) = 1250Hz. These spectra were combined stochastically to represent an experimental acquisition, and the metabolites were able to vary in concentration from 0mmol to 10mmol. A total of 40,100 simulated spectra were generated. Next, non-uniform under-sampling was applied to the indirect dimension to reduce the theoretical acquisition time from 20 minutes to 15, 10, and 5 minutes for three under-sampling schemes. A deep learning model, the densely connected U-Net (D-UNet), was used to learn how to reconstruct the 40,000 training images, and was tested on the remaining 100 images. This model was compared to the standard compressed sensing L1 (non-linear, iterative) reconstruction.
Results: For training, the model achieved a mean squared error (MSE) of 3 X 10â?»â?µ. The D-UNet achieved normalized MSE% of 3.88%, 1.7%, and 3.27% for the 15, 10, and 5 minute acquisitions, respectively. The L1 reconstruction, on the other hand, achieved 1.95%, 6.81%, and 20.8% normalized MSE%. Qualitative comparisons show large artifacts in the L1 reconstructions that are not present in the D-UNet reconstructions.
Conclusion: Deep learning can be used to greatly improve data acquisition time for the LCOSY experiment, as demonstrated with our results. This makes the 2D LCOSY approach more accessible for in vivo chemical quantitation.