MRI basics

MRI basics

Textbooks

  •  MRI From Picture to Proton 
  •  Handbook fMRI analysis 
  •  Functional Magnetic Resonance Imaging (2nd Edition) 
  •  Handbook of Pulse Sequences 
  •  Introduction to functional magnetic resonance imaging 
  •  MRI basic principles and applications 
  •  Magnetic resonance imaging: Physical principles and sequence design 
  •  mri : Physics 
  •  fMRI for dummies 
  •  Principles of fMRI 

Videos

  •  The Insane Engineering of MRI Machines 
  •  MRI Physics overview 

Courses

  •  https://www.coursera.org/learn/functional-mri 

Practical: Types of scans

  • Choose a magnetic field strength (3T, 7T)
  • Choose an RF coil (32-channel, 64-channel)
  • Different coils have different tSNR performance and different spatial profiles. One can acquire some test phantom data to see how different coils fare.
  • Shimming
  • The scanner can automatically (or the user can try to do this manually) try to adjust the strength of the gradient hardware to try to homogenize the imperfections in the constancy of the magnetic field
  • If you want to exert control, you can even prescribe a "shim box" to tell the scanner where to optimize for.
  • "Anatomy" - primarily T1-weighted volumes, but also T2-weighted volumes. Less common: SWI, angiograms
  • "Diffusion" - arguably, this is anatomy, but it's considered a distinct type of data. These volumes are typically T2-weighted but with diffusion weighting.
  • "Functional" - primarily T2*-weighted (i.e. BOLD contrast) volumes
  • "Fieldmaps"
  • Conventional fieldmaps - GRE images at two different echo times
  • Blipped up-down fieldmaps - spin-echo EPI acquisition but with reversed phase-encode directions across the two volumes
  • "Localizer/scout" - quick scan just to see where the brain is (so that you can prescribe slices)

Practical: Typical scanning plans

  • One flavor for fMRI data collection: in one scanning visit, you'll get anatomy (typically: T1), fMRI data (broken up into runs that last 3-10 min each). Other things that might get acquired:
  • Functional localizer experiment (in addition to the main experiment)
  • Fieldmaps (typically just one, but cvnlab likes to do several)
  • Resting-state (5-15 min) (in addition to the main experiment)
  • Downsides: it might be that there's not enough time to collect everything you want!
  • Some experiments plan to collect fMRI data across multiple scanning visits
  • Downsides:
  • time and money (there is overhead to each visit)
  • fMRI data might change across days (e.g., cognitive variability, physiological state variability, data from different sessions also might be harder to align, the absolute MR intensity can and will be different across days)
  • note that you may specifically want to balance your experimental manipulations across days (i.e., make sure to repeat a given condition on each day)
  • Sometimes people really want multimodal data, in which serious time is dedicated to different types of data: T1, T2, diffusion, SWI, etc.

Practical: Choices of pulse sequence parameters

  • Typically, neuroscientists will optimize parameters for their fMRI sequence (but also for anatomy to some extent)
  • fMRI
  • Voxel size (spatial resolution). Typical range: 2mm - 3mm.
  • Coverage (number of slices). Typical range: whole-brain or 1/2 or 1/3 brain.
  • In terms of whole-brain axial slice coverage, HCP7TRET: 85 slices x 1.6 = 136 mm, NSD: 84 slices x 1.8 = 151.2 mm, Kay eLife 2017: 58 slices x 2.5 = 145 mm.
  • TR (temporal resolution). Typical range: 1.5 s - 2.5 s.
  • The RF pulse takes time; slice acquisition takes time; multiple slices takes time; high-res EPI data takes extra time
  • There is some reason that going too fast will reduce the Ernst-angle and cause relatively low amounts of raw MR signal strength; but this is a relatively minor concern
  • Make your TR as low as possible always (ask the scanner console: how low go you go before triggering other changes?)
  • Parallel imaging (GRAPPA, IPAT) acceleration - note that acceleration always comes at SNR costs. Typical range: 1x - 2x (maybe 3x).
  • Simultaneous multi-slice (multiband) acceleration. Typical range: None or 2x-5x.
  • TE. Depends on field strength
  • For 3T, typical range: 30-35 ms
  • For 7T, typical range: ~22-28 ms
  • Partial Fourier. Typical range: 1 (no partial Fourier), 7/8, 6/8
  • Upside: You save time.
  • Downside: You don't actually have as much resolution as you thought you were getting IN the phase-encode dimension.

Basic MRI physics facts you should know

  • Steps to getting an MR signal:
  • The strong static magnetic field created by the scanner (e.g. 3T, 7T) tends to align "spins" of hydrogen protons.
  • There is a transmit RF coil (antenna) that injects (excites) radiofrequency (RF) energy at specific frequencies (Larmor frequency, resonant frequency) will put the object in higher energy state.
  • Then you detect the echo / emitted signal from that object using an receive RF coil (antenna). (The transmit and receive coils might be the same coil.)
  • In this context, what you are measuring is basically proton density.
  • MR contrast
  • Larmor frequency
  • spins in phase vs. spins out of phase
  • (exponential) signal decay
  • T1 - longitudinal relaxation constant
  • T2 - transverse relaxation constant
  • T2* - apparent transverse relaxation constant (apparent because the dephasing can be influenced by macroscopic inhomogeneities)
  • Note that T1 and T2 constants might change drastically over the course of development (first 9 months in newborns, MRI scans look really different!)
  • [T1/T2/T2*]-weighted image - it's primarily the [T1/T2/T2*] of the tissues that are giving rise to differential signal magnitudes
  • Note that "T1 images" and "T2 images" are images of the actual quantitative time constants whereas "T1-weighted images" and "T2-weighted images" imply only that the signal intensity is sensitive to the underlying time constants.
  • TR - repetition time
  • TE - echo time
  • flip angle - angle over which the RF pulse "tips" the longitudinal magnetization
  • readout window - the length of time over which you acquire your MR signal (for a given slice)
  • Note that MRI pulse sequences come in huge varieties, and these sequences can set the TR, TE, etc. at many different points in time and different contexts. So be careful.
  • Intuitions for how to set up TR, TE, flip angle to achieve different tissue weightings
  • short TR, short TE => T1-weighting (slow T1 constants don't have time to recover)
  • long TR, short TE => PD-weighting (full recovery, get total signal)
  • long TR, long TE => T2-weighting (fast T2 dephasing leads to weak signals)
  • explain MPRAGE T1 inversion contrast?
  • Note that higher magnetic field strengths (e.g. 7T) will change tissue constants. In particular, T2 and T2* decrease substantially at high field...
  • Getting a picture or image (i.e. image encoding)
  • 2D sequence: measure individual slices at a time
  • 3D sequence: measure the whole 3D object simultaneously (this will cause blurring over the time it takes to acquire an entire volume)
  • Echo planar imaging (EPI) - most commonly used pulse sequence for fMRI imaging
  • "slice selection" → use gradients to enforce that the RF pulse only perturbs spins within a specific desired slice
  • "image/spatial encoding" → measure different points in k-space
  • k-space
  • phase-encode dimension - this dimension is encoded with "phase encoding" and takes longer to traverse
  • frequency-encode dimension - this dimension is rapidly acquired/traversed (typically with a zig-zag pattern)
  • each little squiggle in the signal is referred to as an "echo"
  • TE for EPI corresponds to when you sample the center of k-space (assuming a Cartesian trajectory). However, note that the entire image readout takes a relatively long time; hence, different points in k-space experience different TEs. This phenomenon leads to what is called T2*-blurring.
  • interleaved - slices are often not acquired in sequential order but interleaved (e.g. first the odd slices, then the even slices)
  • pulse sequence diagram
  • slices
  • field of view - the spatial extent over which a slice is intended to image
  • matrix size - the number of voxels along the frequency encode and phase encode dimensions. For example, the matrix size might be 64 x 64 or 80 x 80.
  • Ernst angle. In EPI, the volume is repeatedly excited (once every TR). Typically, the longitudinal relaxation is not fully recovered by the time of the next excitation. Hence, the signal strength is not fully maximized and there is the concept of a steady-state signal that is eventually achieved (after, say, about 2-3 repeated excitations). The Ernst angle is the flip angle that achieves maximal steady state signal strength. Ernst angle = acos(exp(-TR/T1))/pi*180.
  • What is the BOLD signal?
  • Blood oxygenation-level dependent signal
  • The core phenomenon: deoxygenated hemoglobin vs. oxygenated hemoglobin have different magnetic properties — the deoxygenated blood will tend to cause more disturbances of the magnetic field and lead to a lower net signal (due to spin dephasing)
  • CBF (cerebral blood flow), CBV (cerebral blood volume), CMRO2 (oxygen metabolism rate) → BOLD reflects the combination of all of these quantities!
  • T2*-weighted sequences are sensitive to the BOLD effect

MRI data problems and data quality

  • See  🎩⁠⁠MRI data quality⁠ .
  • How is MRI data sub-optimal?
  • We want all the resolution, but we can't get it. Why not?
  • RF pulses and image readout takes time... meanwhile, the signal is decaying.
  • SNR is always a limitation. (If you push spatial resolution, SNR will suffer.)
  • SNR
  • A quick and easy metric of SNR is "temporal signal-to-noise ratio (tSNR)" which basically is the mean across volumes divided by the std dev across volumes
  • Makes most sense for a phantom since we don't expect real signal fluctuations (but for humans, there are certainly resting-state fluctuations that are real signals in the data)!
  • fMRI
  • spatial distortion - the images we get aren't fully spatially accurate but have local distortions in them. this is annoying for localizing your signals properly.
  • shortening the image readout will lessen distortion
  • parallel imaging will reduce distortion (since it shortens the readout)
  • signal dropout — orbitofrontal and medial temporal are the two main locations that exhibit severe dropout.
  • thinner slices will tend to reduce dropout
  • shorter TE will tend to reduce dropout (but you may lose BOLD sensitivity)
  • improving the overall magnetic field shim can help reduce dropout to some degree
  • Non-quantitative
  • It would be nice if the numbers we got actually reflect physical units.
  • What the heck is BOLD, biologically speaking?
  • Relationship to spiking, LFP, electrophysiology, etc. is complex