Signal and Noise in fMRI John VanMeter, Ph.D. Center for Functional and Molecular Imaging Georgetown University Medical Center
Outline Definition of SNR and CNR in context of anatomic imaging Definition of functional SNR Sources of noise in MRI Source of noise in fMRI Changes in MRI SNR and functional SNR with increased magnetic field strength
MRI Signal and Noise Signal is primarily dependent on number of protons in the voxel Noise can come from RF energy leaking into the scanner room, random fluctuations in electrical current, etc. The body creates noise in the MR signal via changes in current in the body producing small changes in the magnet field; breathing can change homogeneity
Measuring MRI Signal-to-Noise Ratio (SNR) Signal is the intensity (brightness) of one or more pixels in the object of interest. Noise is the intensity of one or more pixels in the ‘air’ (i.e. outside the object of interest). SNR = Signal (low SNR = grainy, fuzzy images) Noise Fundamental measure of image quality
MRI SNR – Example 1 S N S = 700 N = 20 SNR = 700 / 20 = 35
MRI SNR – Example 2 S N S = 300 N = 50 SNR = 300 / 50 = 6
MRI SNR - Side-by-Side SNR = 35SNR = 6
SNR in Terms of fMRI MRI SNR is not the most important issue with regard to functional MRI Functional SNR is contingent on ability to detect changes in BOLD signal between conditions (across time) Underlying MRI SNR still important in terms of providing base for signal in functional SNR but several other factors affect signal and noise in fMRI data
Affect of MRI SNR on Functional SNR Increase in MRI signal due to BOLD affect “rides on top” of signal of in MRI scan Imagine 2% increase in signal between these two fMRI scans In which image will the 2% change be more detectable?
Changes in BOLD Signal are Small Visual and sensorimotor areas percent change might be as high 5% For most other cortical areas expected percent change is on the order of 1-3%
Measuring Percent Signal Change
MRI Contrast-to-Noise Ratio (CNR) Measure of separation in terms of average intensity between two tissues of interest Defined as difference between the SNR of the two tissues (A & B): CNR = Signal A – Signal B Noise
MRI CNR – Example 1 S W = 700, S G = 200 N = 20 CNR WG = (700 – 200) / 20 = 25 SWSW N SGSG
MRI CNR – Example 2 S W = 200, S G = 100 N = 50 CNR WG = (200 – 100) / 50 = 2 SWSW N SGSG
MRI CNR Side-by-Side SWSW N SGSG SWSW N SGSG CNR WG = 35CNR WG = 6
Functional CNR vs Functional SNR Generally CNR is unimportant in fMRI as there is little contrast between tissues Some researchers refer to difference between “On” and “Off” as dynamic CNR or functional CNR Probably more accurate to refer to ability to detect changes related to activity as functional SNR
Functional SNR is a dependent on differences in signal across time Ability to distinguish differences between different conditions - effect size
Differences Between Two Conditions Typically compare BOLD signal in the same area under different conditions Example fusiform face area; responds to both faces and tools but about 0.2% more to faces
Sources of Noise in fMRI Data System noise –Thermal noise –Signal drift Subject dependent noise Physiological noise Variability in BOLD response Variability across sessions within subject Variability across subjects
Thermal Noise Intrinsic noise due to thermal motion of electrons –In subject –In RF equipment Increases with temperature - atoms move faster; more collisions; greater loss of energy Unfortunately increases with field strength approximately linearly Effects limited to temporal fluctuations and is equally likely to add or subtract thus roughly Gaussian (i.e. normally) distributed
Signal Drift Across Time Magnetic field has slight drifts in strength over time produces drift in signal Gradually, over time the MRI signal in a voxel drifts This drift can vary from one voxel to the next both in degree and direction!
Signal Drift
Affect of Signal Drift
Effect of Nonlinear Drifts
Physiological Noise Subject movement during scan –Single largest source of noise in fMRI data –Extremely problematic if motion is timed with task –Makes studies with overt speech during the scan quite difficult –Motion more problematic across time points
Subject Motion
Pulsatile Motion of Brain Influx of blood into brain induces movement especially around base of brain - why there? Short TR’s can also pick-up noise due to respiration (TR<2500ms) and cardiac (TR<500ms) cycle
Map showing standard deviation of intensity over time Two sources of noise evident Why do edges of brain show large effect? Often referred to as “ringing”
Power Spectrum
Other Sources of Physiological Noise Change in CO 2 - hyperventilation produces change in O 2 content of blood; blood flow increases to compensate Drug affects - antihistamines, etc Smokers vs. Non-smokers –Hypoactivation on attentional task after abstaining for 1hr reversed following nicotine patch (Lawrence et al, 2002)
Genetic Based Differences ApoE risk factor for Alzheimer’s disease Study of non- symptomatic carriers Reduced activation in hippocampus on a memory task for high risk carriers (AS Fleisher, et al, Neurobiology of Aging, 2008)
Noise from Neural Activity Not of Interest Eye movements - results in activation of the frontal eye-fields Noise of the scanner - activates auditory cortices –Usually not a problem as noise common to both conditions –Auditory experiments difficult though Other thoughts - what’s for dinner, going over a to-do list, wondering what the experiment is testing (grad students), etc
Behavioral and Cognitive Variability Passive tasks are prone to drift in subject attention and/or arousal –Difficult to identify performance on tasks and compare across subjects Tasks with responses can lead to variations in reaction/response time –Speed-accuracy trade-off Task strategies used can differ Task difficulty especially between groups of subjects very problematic
Inter-Subject Variability
Inter-Session Variability
Intra-Session Variability
99-Scanning Sessions Same subject participated in 99 identical scanning sessions 33 each for motor task, visual task, and a cognitive task Everything kept exactly the same Considerable variability was observed
33 Motor Sessions McGonigle, et al., Neuroimage, 2000
33 Cognitive Sessions
Strategies for Dealing with Noise & Improving Signal MRI Center Steps –Measure stability of signal over time –Ensure stability of equipment –Eliminate RF-noise Researcher –Formalize instructions (use scripts) –Train subjects ahead of time –Instruct subjects to use same strategy –Stress importance of staying still, focus, etc. Use better post-processing techniques Increase field strength
Post-processing Pre Post Smith, et al., Human Brain Mapping, 2005
Signal Averaging Averaging across multiple trials greatly helps to improve SNR Each graph shows 20 traces of 1 trial, average of 4 trials, average of 9 trials, etc
Increasing MRI Signal with Stronger Magnets Increase magnetic field strength –Plus: more protons pulled into alignment thus greater net magnetization resulting in increased MRI signal –Minus: shortens T2* resulting in larger spatial distortions with gradient echo sequences Requires larger RF pulses thus SAR goes up (why?)
Susceptibility Distortion Increases with Field Strength 1.5T 4.0T
Rules of Thumb Quadratic increase in MRI signal with increase in field strength Thermal noise scales linearly with field strength Raw MRI SNR thus only scales linearly What about functional SNR?
Functional SNR Linearly Increases with Field Strength?
Functional SNR vs Field Strength MRI signal goes up quadratically Thermal noise goes up linearly Physiological noise goes up quadratically Eventually functional SNR expected to plateau
Upsides to Field Strength for Functional SNR Increase in number of voxels activated and presumably detectability T2* of blood much shorter thus signal drops off in larger vessels –Linear increase in large vessels –Quadratic increase in small vessels –Thus, spatial specificity increases