Clean BOLD for better stats
What fMRI noise and what to do about it?
Two cool papers came out in neuroimage: Noise contributions to the fMRI signal: An overview from Thomas Liu and Methods For Cleaning The BOLD fMRI Signal from César Caballero-Gaudes and Richard C. Reynolds. One is describing noise and how to deal with it, the other discusses processing pipelines. Of course current solutions may not all fit with what Liu is presenting because tools tend to lag behind. Let's dig in !
In Lius' paper there are few interesting measures to discuss.
The 1st one is temporal SNR. This is a classic in the fMRI literature and is simply the mean signal / std over time. There are however variations on the way it is computed: in-mask, per tissue, or even as proposed in the paper separating background, BOLD (presumably derived from gray matter) and non BOLD (presumably derived from CSF).
Another measure we could all acquire (but I personally never seen this used) is background noise, typically made up of thermal noise and RF interferences. Easy to measure: acquire data setting the flip angle to 0! In addition, to check the scanner doesn't have too many of these RF interferences, we can compute tSNR as a function of ROI size - and this should increase as the square root of the number of voxels.
The cardiac cycle is a major source of variance, we know it's roughly peaking at 1Hz. The other major issue is respiration and this is peaking at about 0.25Hz and fluctuations in CO2 (a vasodilatator) at a much lower frequency around 0.02Hz. This last one is thus the real problem for rs-fMRI since network localization seems to many occur in the 0.01 - 0.1Hz band. That doesn't mean we should not care about the others, because the TR is usually long (say 2sec) and therefore the cardiac and respiratory noise will appear at a much lower frequency due to aliasing. For task based people like me, we want design with conditions that are well above those frequencies, an old fashion rule that we don't see much these days (because people use defaults) is to filter at twice the lowest frequency among conditions (or twice the maximum average time occurrence).
An interesting point raised in these papers (referencing others) is that noise interacts / influences the BOLD measurements and filtering is not always 100% efficient. One option, is to use additional regressors in your GLM. What regressors to use? rs-fMRI tells us that we can derived relatively good estimates from white matter and csf. Typically, taking the strongest components from the PCA across voxels rather than just the average since the global signal derived from this procedure also contains interesting signals that we don't want to regress out.The other option is to model these changes using e.g. RETROICOR or DRIFTER - the later one having the advantage that it can even estimate from the data the cardiac and respiratory oscillations, while in the former one must have external measurements.
More measure of noise: although not common, plotting the power spectrum in each tissue class after data have been processed/cleaned can easily tell you if the data processing pipeline did a good job. Not only you can see if you still have physiological noise but also doing so on residuals from a GLM will tell you if they are iid: if all goes well, the power spectrum is flat.
The other big problem is motion. In a previous post, replicating work done at Berkeley by practiCal fMRI, I looked at minimizing motion in high resolution fMRI (1.5mm isotropic). Whatever data acquisition (prospective MoCo) or data processing technique we can come up with, I think it will always be best to minimize as much as possible using some sort of head motion restrain (simple padding is not enough).
For GLM approaches, it is clear that absolute motion is not the biggest issue (6 motion parameters). One real problem is motion within a scan, leading to spin history effect, and this cannot be corrected properly. Here ICA seems to be doing a decent job at dealing with both type of motion and might be better suited, assuming automated methods to label components works well. The other big problem is movement between scans and thus using the derivative (12 motion parameters) and/or motion scrubbing seems like a good idea. Scrubbing is efficiently implemented using nulling regressors (0 everywhere but the scan of interest). One problem is how to identify bad scans. Using displacement alone might not be enough, and looking at temporal outliers (e.g. DVARS) can be more efficient, especially because artefacts due to motion can last longer than a single scan that would be identified by looking at displacement alone. Scrubbing is however not recommended if instead one uses regressors derived from white matter or csf, like CompCor. In addition, in an approach like CompCor (there a few proposals out there) regressors from temporally varying outlier voxels are also generated, thus accounting for spikes and other artefacts. The problem here is to define appropriate masks to derive regressors - if the mask include voxels of interest we might regress out the signal we are looking for. Caballero-Gaudes and Reynolds propose to orthogonalize such regressors with regards to the task regressors to avoid removing signal of interest. I think this is not a good solution, because if the signal if confunded in the noise, orthogonalization won't remove noise, since it is captured by task regressors. Yes, that means less signal because of extra regressors are not orthogonal, but maybe better signal because not contaminated by noise (better as in more reliable).
Another aspect of motion, is how it interacts with the magnetic field that is not as homogenous as we would like and field maps should be used to (partly) compensate distortions that occur (note the distortion due to motion can include motion due to respiration).
Caballero-Gaudes and Reynolds suggest a series of useful preprocessing steps: despiking, physiological noise correction, realignment with distortion correction for task based fMRI. For rs-fMRI the proposed solution is a GLM with filtering and motion regressors and scrubbing, all in one to avoid artefacts due to filtering.
Thomas Liu (2016) Noise contributions to the fMRI signal: An overview NeuroImage, 343, 141-151.
César Caballero-Gaudes and Richard C. Reynolds (2016). Methods For Cleaning The BOLD fMRI Signal. NeuroImage, in press
Measuring fMRI noise
In Lius' paper there are few interesting measures to discuss.
The 1st one is temporal SNR. This is a classic in the fMRI literature and is simply the mean signal / std over time. There are however variations on the way it is computed: in-mask, per tissue, or even as proposed in the paper separating background, BOLD (presumably derived from gray matter) and non BOLD (presumably derived from CSF).
Another measure we could all acquire (but I personally never seen this used) is background noise, typically made up of thermal noise and RF interferences. Easy to measure: acquire data setting the flip angle to 0! In addition, to check the scanner doesn't have too many of these RF interferences, we can compute tSNR as a function of ROI size - and this should increase as the square root of the number of voxels.
Physiological noise sources
The cardiac cycle is a major source of variance, we know it's roughly peaking at 1Hz. The other major issue is respiration and this is peaking at about 0.25Hz and fluctuations in CO2 (a vasodilatator) at a much lower frequency around 0.02Hz. This last one is thus the real problem for rs-fMRI since network localization seems to many occur in the 0.01 - 0.1Hz band. That doesn't mean we should not care about the others, because the TR is usually long (say 2sec) and therefore the cardiac and respiratory noise will appear at a much lower frequency due to aliasing. For task based people like me, we want design with conditions that are well above those frequencies, an old fashion rule that we don't see much these days (because people use defaults) is to filter at twice the lowest frequency among conditions (or twice the maximum average time occurrence).
An interesting point raised in these papers (referencing others) is that noise interacts / influences the BOLD measurements and filtering is not always 100% efficient. One option, is to use additional regressors in your GLM. What regressors to use? rs-fMRI tells us that we can derived relatively good estimates from white matter and csf. Typically, taking the strongest components from the PCA across voxels rather than just the average since the global signal derived from this procedure also contains interesting signals that we don't want to regress out.The other option is to model these changes using e.g. RETROICOR or DRIFTER - the later one having the advantage that it can even estimate from the data the cardiac and respiratory oscillations, while in the former one must have external measurements.
More measure of noise: although not common, plotting the power spectrum in each tissue class after data have been processed/cleaned can easily tell you if the data processing pipeline did a good job. Not only you can see if you still have physiological noise but also doing so on residuals from a GLM will tell you if they are iid: if all goes well, the power spectrum is flat.
Motion
The other big problem is motion. In a previous post, replicating work done at Berkeley by practiCal fMRI, I looked at minimizing motion in high resolution fMRI (1.5mm isotropic). Whatever data acquisition (prospective MoCo) or data processing technique we can come up with, I think it will always be best to minimize as much as possible using some sort of head motion restrain (simple padding is not enough).
For GLM approaches, it is clear that absolute motion is not the biggest issue (6 motion parameters). One real problem is motion within a scan, leading to spin history effect, and this cannot be corrected properly. Here ICA seems to be doing a decent job at dealing with both type of motion and might be better suited, assuming automated methods to label components works well. The other big problem is movement between scans and thus using the derivative (12 motion parameters) and/or motion scrubbing seems like a good idea. Scrubbing is efficiently implemented using nulling regressors (0 everywhere but the scan of interest). One problem is how to identify bad scans. Using displacement alone might not be enough, and looking at temporal outliers (e.g. DVARS) can be more efficient, especially because artefacts due to motion can last longer than a single scan that would be identified by looking at displacement alone. Scrubbing is however not recommended if instead one uses regressors derived from white matter or csf, like CompCor. In addition, in an approach like CompCor (there a few proposals out there) regressors from temporally varying outlier voxels are also generated, thus accounting for spikes and other artefacts. The problem here is to define appropriate masks to derive regressors - if the mask include voxels of interest we might regress out the signal we are looking for. Caballero-Gaudes and Reynolds propose to orthogonalize such regressors with regards to the task regressors to avoid removing signal of interest. I think this is not a good solution, because if the signal if confunded in the noise, orthogonalization won't remove noise, since it is captured by task regressors. Yes, that means less signal because of extra regressors are not orthogonal, but maybe better signal because not contaminated by noise (better as in more reliable).
Another aspect of motion, is how it interacts with the magnetic field that is not as homogenous as we would like and field maps should be used to (partly) compensate distortions that occur (note the distortion due to motion can include motion due to respiration).
A proposed workflow
Caballero-Gaudes and Reynolds suggest a series of useful preprocessing steps: despiking, physiological noise correction, realignment with distortion correction for task based fMRI. For rs-fMRI the proposed solution is a GLM with filtering and motion regressors and scrubbing, all in one to avoid artefacts due to filtering.
References
Thomas Liu (2016) Noise contributions to the fMRI signal: An overview NeuroImage, 343, 141-151.
César Caballero-Gaudes and Richard C. Reynolds (2016). Methods For Cleaning The BOLD fMRI Signal. NeuroImage, in press
Comments
Post a Comment