Class 5: Signal, Noise, and the fMRI data preprocessing fMRI: theory and practice Spring 2010
The more steps done at once, the greater the chance of problems The Black Box The danger of automated processing and fancy images is that you can get blobs without every really looking at the real data The more steps done at once, the greater the chance of problems Big Black Box of automated software Raw Data Pretty pictures
Know Thy Data Look at raw functional images Look at the movies Where are the artifacts and distortions? How well do the functionals and anatomicals correspond Look at the movies Is there any evidence of head motion? Is there any evidence of scanner artifacts (e.g., spikes) Look at the time courses Is there anything unexpected (e.g., abrupt signal changes at the start of the run)? What do the time courses look like in the unactivatable areas (ventricles, white matter, outside head)? Look at individual subjects Double check effects of various transformations Make sure left and right didn’t get reversed Make sure functionals line up well with anatomicals following all transformations Think as you go. Investigate suspicious patterns.
Hardware Malfunctions Sample Artifacts Ghosts Hardware Malfunctions Metallic Objects (e.g., hair tie) Spikes Longitudinal saturation effect
Calculating Raw Signal:Noise Ratio Pick a region of interest (ROI) outside the brain free from artifacts (no ghosts, susceptibility artifacts). Find mean () and standard deviation (SD). Pick an ROI inside the brain in the area you care about. Find and SD. e.g., =4, SD=2.1 SNR = brain/ outside = 200/4 = 50 [Alternatively SNR = brain/ SDoutside = 200/2.1 = 95 (should be 1/1.91 of above because /SD ~ 1.91)] When citing SNR, state which denominator you used. e.g., = 200 Head coil should have SNR > 50:1 Surface coil should have SNR > 100:1 Source: Joe Gati, personal communication
Calculating Functional SNR
What affects SNR? Physical factors SOLUTION & TRADEOFF Thermal Noise (body & system) Inherent – can’t change Magnet Strength e.g. 1.5T 4T gives 2-4X increase in SNR Use higher field magnet additional cost and maintenance physiological noise may increase Coil e.g., head surface coil gives ~2+X increase in SNR Use surface coil Lose other brain areas Lose homogeneity Voxel size e.g., doubling slice thickness increases SNR by root-2 Use larger voxel size Lose resolution Sampling time Longer scan sessions additional time, money and subject discomfort Source: Doug Noll’s online tutorial
Coils Head coil Surface coil homogenous signal moderate SNR highest signal at hotspot high SNR at hotspot Source: Joe Gati
Bigger is better… to a point Voxel size Bigger is better… to a point Increasing voxel size signals summate, noise cancels out “Partial voluming”: If tissue is of different types, then increasing voxel size waters down differences e.g., gray and white matter in an anatomical e.g., activated and unactivated tissue in a functional
Sampling Time More samples More confidence effects are real
What affects SNR? Physiological factors SOLUTION & TRADEOFF Head (and body) motion Use experienced or well-warned subjects limits useable subjects Use head-restraint system possible subject discomfort Post-processing correction often incompletely effective 2nd order effects can introduce other artifacts Single trials to avoid body motion Cardiac and respiratory noise Monitor and compensate hassle Low frequency noise Use smart design Perform post-processing filtering BOLD noise (neural and vascular fluctuations) Use many trials to average out variability Behavioral variations Use well-controlled paradigm Source: Doug Noll’s online tutorial
Effects of field strength
Showing visual columns in 7T
But bigger magnet has its downside…
Sources of Noise Physical noise “Blame the magnet, the physicist, or the laws of physics” Physiological noise “Blame the subject”
Freq distribution of physiological noise
A Map of Noise voxels with high variability shown in white
Linear Drift (or scanner drift)
Distribution of physiological noise
Field Strength Although Raw SNR goes up with field strength, so does thermal and physiological noise Thus there are diminishing returns for increases in field strength
Data Preprocessing Options 1. artifact screening ensure the data is free from scanner and subject artifacts done by eyeballing and manual correction 2. slice scan time correction correct for sampling of different slices at different times 3. motion correction 4. spatial filtering smooth the spatial data 5. temporal filtering remove low frequency drifts (e.g., linear trends) remove high frequency noise (not recommended because it increases temporal autocorrelation and artificially inflates statistics) 6. spatial normalization put data in standard space (Talairach or MNI Space)
2. slice-scan time correction
3. Motion-induced intensity Changes Slide modified from Duke course
Motion Spurious Activation at Edges lateral motion in x direction motion in z direction (e.g., padding sinks) brain position time1 time2 time 1 > time 2 time 1 < time 2 stat map
Spurious Activation at Edges
Motion Correction Algorithms pitch roll yaw z translation y translation x translation Most algorithms assume a rigid body (i.e., that brain doesn’t deform with movement) Align each volume of the brain to a target volume using six parameters: three translations and three rotations Target volume: the functional volume that is closest in time to the anatomical image
Head Motion: relatively good
… and catastrophically bad Slide from Duke course
Problems with Motion Correction lose information from top and bottom of image possible solution: prospective motion correction calculate motion prior to volume collection and change slice plan accordingly we’re missing data here we have extra data here Time 1 Time 2
Different motions; different effects Drift within run Movement between runs Uncorrelated abrupt movement within run Correlated abrupt movement within a run Motion correction Spurious activations okay, corrected by LTR okay minor problem huge problem can reduce problems Increased residuals problem can reduce problems; may be improved by including motion parameters as predictors of no interest Regions move minor-major problem depending on size of movement can reduce problems; if algorithm is fooled by physics artifacts, problem can be made worse by MC Physics artifacts not such a problem because effects are gradual can’t fix problem; may be misled by artifacts
The Fridge Rule When it doubt, throw it out!
(more comfortable than it sounds!) Head Restraint Vacuum Pack Head Vise (more comfortable than it sounds!) Bite Bar Thermoplastic mask Often a whack of foam padding works as well as anything
Even the mock scanner…
Prevention is the Best Remedy Tell your subjects how to be good subjects “Don’t move” is too vague Make sure the subject is comfy going in avoid “princess and the pea” phenomenon Emphasize importance of not moving at all during beeping do not change posture if possible, do not swallow do not change mouth position do not tense up at start of scan Discourage any movements that would displace the head between scans Do not use compressible head support
4. Spatial Smoothing Application of Gaussian kernel Usually expressed in #mm FWHM “Full Width – Half Maximum” Typically ~2 times voxel size Slide from Duke course
Reduction of false-positive rate by spatial smoothing
Effects of Spatial Smoothing on Activity Unsmoothed Data Smoothed Data (kernel width 5 voxels) Slide from Duke course
Should you spatially smooth? Advantages Increases Signal to Noise Ratio (SNR) Matched Filter Theorem: Maximum increase in SNR by filter with same shape/size as signal Reduces number of comparisons Allows application of Gaussian Field Theory May improve comparisons across subjects Signal may be spread widely across cortex, due to intersubject variability Disadvantages Reduces spatial resolution Challenging to smooth accurately if size/shape of signal is not known Slide from Duke course
5. Time Course Filtering
Low and High Frequency Noise Source: Smith chapter in Functional MRI: An Introduction to Methods
Preprocessing Options Before LTR: After LTR:
Preprocessing Options High pass filter pass the high frequencies, block the low frequencies a linear trend is really just a very very low frequency so LTR may not be strictly necessary if HP filtering is performed (though it doesn’t hurt) Before High-pass After High-pass linear drift ~1/2 cycle/time course ~2 cycles/time course
Preprocessing Options Gaussian filtering each time point gets averaged with adjacent time points has the effect of being a low pass filter passes the low frequencies, blocks the high frequencies You better know it clearly what you are doing Before Gaussian (Low Pass) filtering After Gaussian (Low Pass) filtering
Take home Messages Look at your data Work with your physicist to minimize physical noise Design your experiments to minimize physiological noise Motion is the worst problem: When in doubt, triple-check Preprocessing is not always a “one size fits all” exercise