5. The spread of the boxes correspond … In vivo application Similar to the procedure used with simulated data, in vivo
spectral data (N = 193) were demeaned and rank reduced using singular value decomposition, to 20 components, before multirun ICA. The extracted ICs were automatically paired with LCModel basis and corresponding BIBF 1120 mouse weights were also estimated. Table 2 lists those select pairs with significant spectral correlations, and captures both spectral and weights correlations. While ICs resembling Inhibitors,research,lifescience,medical the m-Ins signal and the singlet resonances of NAA, NAAG, Cr, PCh, and s-Ins were readily identified, no ICs resembling resonances from Asp, Glu, Gln, and GABA were discerned. The table also captures how the ICA and LCModel Inhibitors,research,lifescience,medical estimates relate to the fractional tissue volume in the spectroscopic voxel. More the tissue fraction, more signal is detected and the estimates are larger. Therefore, without normalization, the estimates show similar, positive correlations with tissue volumes; the correlations are weak possibly due to the lack of perfect spatial overlap between Inhibitors,research,lifescience,medical metabolite and water
volumes. However, when normalized neither set of estimates correlates with tissue volumes, as expected. Table 2 Results from ICA analysis of 193 spectra in vivo data: Components identified based on spectral correlation with matching LCModel spectra are shown. Inhibitors,research,lifescience,medical The correlations between the LCModel and ICA estimates (weights), both NAA normalized, are appreciable … Figure 7 shows results from ICA analysis of in vivo data, in the absence any ground truth, plotted against LCModel Inhibitors,research,lifescience,medical references. The components with significant spectral correlations are overlaid on the matching real part of the paired LCModel basis spectrum; spectra plotted are demeaned and intensity normalized. Notice the components substantially overlap paired basis spectra at the
major peaks, with some differences apparent around the baseline, attributed to covarying resonances; for example, the peaks around 2.4 ppm of NAA-like component seem to arise from Glu, based on Pearson correlation in the spectral subspace (r = 0.612). Resonances such as those from Asp, GABA, or Gln are not readily discerned from in vivo data and therefore not presented. Ketanserin Also shown below each set of spectra are the scatter plots of the ICA estimates (weights), plotted against LCModel estimates, both expressed as a ratio with NAA; least squares fit lines for the scatter plots are also shown. As NAA is the reference metabolite, its scatter plot is not constructed; instead, we present a scatter plot between the weights of NAA component and the peak value of the spectral input to ICA.