SPINS: Review of QC metrics
from eddy and MRTrix
eddy and MRTrixcode written: 2020-01-17
last ran: 2020-01-17
website: http://rpubs.com/navona/SPINS_DWI_QCeddyMRTrix
Description. This script evaluates the relationship between quality metrics from eddy quad (n=5) and MRTrix, namely the calculation of spherical harmonic noise residuals (n=1). First, we examine correlations between the metrics. Then, we perform a PCA and examine loadings. We decide if PC1, across these n=6 QC metrics, should be used as a regressor indicating scan quality for subsequent analyses.
Dataset. This analysis includes SPINS DWI data that has not undergone correction for (known) site/scanner effects. We include data from all 422 eligible HC and SSD participants from across all 6 sites, minus n=3 participants (SPN01_CMP_0219, SPN01_CMP_0220, SPN01_ZHH_0009) for whom MRTrix noise residuals failed (I will rerun the pipeline to add these participants). Thus, this analysis includes a total of 419 participants, some of whom will ultimately fail eddy (automated) QC and/or visual QC.
Correlations. As expected, Residual noise shows a negative correlation with Average SNR and Average CNR. However, contrary to expectation, Residual noise is not highly/strongly correlated with Absolute motion or Relative motion, and even shows a small negative correlation with Percent outliers.
PCA. Before performing PCA, the variables were centered and scaled (skewness and the magnitude of the variables influence the resulting PCs), but no other transformation was applied. We see that at least 2 PCs are required to describe a substantial portion of the variance. We also see that Residual noise does not load onto the same factors as the eddy variables.
## Standard deviations (1, .., p=6):
## [1] 1.6506712 1.1752728 0.8885973 0.7475469 0.5506600 0.4923014
##
## Rotation (n x k) = (6 x 6):
## PC1 PC2 PC3 PC4
## Absolute motion 0.4144932 -0.2705421 0.5539088297 0.45371856
## Relative motion 0.4883670 -0.3104089 -0.0246046705 0.24297548
## Percent outliers 0.2300847 -0.5435550 -0.7159339933 -0.05778632
## Average SNR -0.5049828 -0.1788036 -0.0701892172 0.51895562
## Average CNR -0.4871249 -0.3444246 0.0001401093 0.32333271
## Residual noise 0.2108453 0.6200017 -0.4184394699 0.59825091
## PC5 PC6
## Absolute motion 0.49225710 -0.003126194
## Relative motion -0.77798523 0.015507069
## Percent outliers 0.36661840 0.036044256
## Average SNR -0.06822356 0.658882152
## Average CNR -0.08193489 -0.729947652
## Residual noise 0.08151748 -0.177486067## Importance of components:
## PC1 PC2 PC3 PC4 PC5 PC6
## Standard deviation 1.6507 1.1753 0.8886 0.74755 0.55066 0.49230
## Proportion of Variance 0.4541 0.2302 0.1316 0.09314 0.05054 0.04039
## Cumulative Proportion 0.4541 0.6843 0.8159 0.90907 0.95961 1.00000
PCA leaving out MRTrix noise residual. Here, we run the same analysis as above, expect with the Residual noise metric removed.
## Standard deviations (1, .., p=5):
## [1] 1.6274761 1.0412752 0.8378321 0.5542491 0.5078509
##
## Rotation (n x k) = (5 x 5):
## PC1 PC2 PC3 PC4 PC5
## Absolute motion 0.4433073 -0.1054759 0.7541546 -0.46169320 -0.10219433
## Relative motion 0.5137166 -0.2919356 0.1502289 0.78935526 0.07223866
## Percent outliers 0.2637773 -0.7499147 -0.4865120 -0.34678395 -0.10533997
## Average SNR -0.5021182 -0.3531612 0.2771929 0.20831349 -0.70917140
## Average CNR -0.4667684 -0.4653532 0.3084676 0.01020035 0.68579685## Importance of components:
## PC1 PC2 PC3 PC4 PC5
## Standard deviation 1.6275 1.0413 0.8378 0.55425 0.50785
## Proportion of Variance 0.5297 0.2168 0.1404 0.06144 0.05158
## Cumulative Proportion 0.5297 0.7466 0.8870 0.94842 1.00000
Decision. We see that the MRTrix residual seems to be capturing something that isn’t precisely motion. Overall, it appears that PC1 is tracking ‘quality’. In comparison, PC2 is ambiguous. We will include the first PC value across all 6 DWI QC metrics as a regressor for subsequent analyses.