Module 1
Lecture - Introduction
What is fMRI, really?
● Brain imaging has exploded over the last ~30 years.
● Any hospital has an MRI scanner that’s also usable for
cognitive neuroscience research.
● Brain imaging is changing how people think about
themselves; it’s placed our psyches firmly in our heads
Functional Brain Imaging
● fMRI is so popular because it is considered
non-invasive, and provides an appealing balance
between temporal and spatial resolution
● fMRI is evolving: it’s becoming faster and higher
resolution
functional MRI: What is our view on the brain like?
● The signals we measure with fMRI are primarily
dependent on the oxygen level in the blood, they
are BOLD, for Blood Oxygenation Level Dependent
● Since neural activity consumes oxygen, the level of oxygen in the blood will vary as
neurons activate.
● Thus, BOLD is a metabolic measurement.
● Keep in mind: we are measuring not neural activity itself, but a derivative of it. This
means BOLD provides only an indirect view on neural activity!
Blood response: Haemodynamic response
● Slow response to neural activation, takes over
20s to unfold
● Is dependent on a complex cascade of biological,
metabolic responses, among which Cerebral
Blood Flow, and Oxygenation
● Has a (more or less) stereotypical shape:
MRI scanner: A “standard” clinical device, repurposed for cognitive science
● Won the 2003 Nobel prize
● Used to create many different types of images of the living brain for diagnosis, both
different types of anatomical images and different types of functional images
● Different images are created by running different programs, called sequences
1
, ● Different field strengths, measured in Tesla (Earth magnetic field: 3.2 × 10-5 T)
● MR Physics lecture Feb 22nd Nikos Priovoulos
1.5 T 7T 11 T
MRI data: What does MRI data look like?
● It’s 3D: values at {x, y, z} represents the
anatomy
● Elements are called voxels, for volume
element, like picture element for pixels
● Standard anatomy has a resolution of
1mm
● At 7T, the resolution goes to ~0.3mm (27 times smaller,
because cubic relation between voxel size and volume!)
Other MRI data: What a versatile machine!
● Examples of 4D or higher-D images - store not a time-course for each voxel, but more
complex information
fMRI data: What does fMRI data look like?
2
,fMRI data: Data quality
● We are focused on 1. Signal, and 2. Noise - we want more signal, and less noise
● Measure for Scan Quality:
○ tSNR (temporal Signal to Noise Ratio)
μ
○ σ
, where μ is the mean image, and σ is the standard deviation over time
○ Normal values:
■ ~ 0, air, (μ close to 0)
■ > 40, noisy gray matter (more noise, higher σ)
■ > 80, high-signal gray matter (less noise, lower σ)
■ > 100, white matter (almost no noise, very low σ)
fMRI data: fMRI images as data
● Values in a normal fMRI dataset are not expressed in any SI unit, such as meter,
second, Tesla, etc.
● We cannot really compare across scanners, subjects, etc.
● Our workaround measure for BOLD signal strength:
○ Percentage signal change (%sc)
○ , where Ii is the image at timepoint i
○ Normal values:
■ 0.2%-1% (normal fMRI experiment @ 3T),
■ 1%-3% (finger-tapping in primary motor cortex @ 3T)
■ 1%-7% (finger-tapping in primary motor cortex @ 7T)
■ 7%-20% (voxel in a large vein in primary cortex @ 7T, where μ is very
low)
3
, fMRI analysis: How do you get to
results?
● Much more than just a single step
● Experimental design is
paramount:
● What is your goal?
● Many ways of going wrong
Replication Crisis: Many ways of going
wrong?
● With such complex analyses,
room for 1. ’Tinkering’, or 2.
Sloppiness
● We will teach you the newest approaches to reproducible neuroscience:
○ How to save your data
○ How to preprocess your data
○ How to program your data analyses
○ How to report your results
○ How to open-source your data and analyses
“Modern” fMRI: What’s the state of the art? - Naselaris et al, 2021 paper on Perusall
● With higher field strengths, more is possible. For instance, higher resolution means
“laminar” resolution
● Large, open science projects:
○ Human Connectome Project (HCP, 1200 sjs, >5 hours, 3T & 7T)
4
Lecture - Introduction
What is fMRI, really?
● Brain imaging has exploded over the last ~30 years.
● Any hospital has an MRI scanner that’s also usable for
cognitive neuroscience research.
● Brain imaging is changing how people think about
themselves; it’s placed our psyches firmly in our heads
Functional Brain Imaging
● fMRI is so popular because it is considered
non-invasive, and provides an appealing balance
between temporal and spatial resolution
● fMRI is evolving: it’s becoming faster and higher
resolution
functional MRI: What is our view on the brain like?
● The signals we measure with fMRI are primarily
dependent on the oxygen level in the blood, they
are BOLD, for Blood Oxygenation Level Dependent
● Since neural activity consumes oxygen, the level of oxygen in the blood will vary as
neurons activate.
● Thus, BOLD is a metabolic measurement.
● Keep in mind: we are measuring not neural activity itself, but a derivative of it. This
means BOLD provides only an indirect view on neural activity!
Blood response: Haemodynamic response
● Slow response to neural activation, takes over
20s to unfold
● Is dependent on a complex cascade of biological,
metabolic responses, among which Cerebral
Blood Flow, and Oxygenation
● Has a (more or less) stereotypical shape:
MRI scanner: A “standard” clinical device, repurposed for cognitive science
● Won the 2003 Nobel prize
● Used to create many different types of images of the living brain for diagnosis, both
different types of anatomical images and different types of functional images
● Different images are created by running different programs, called sequences
1
, ● Different field strengths, measured in Tesla (Earth magnetic field: 3.2 × 10-5 T)
● MR Physics lecture Feb 22nd Nikos Priovoulos
1.5 T 7T 11 T
MRI data: What does MRI data look like?
● It’s 3D: values at {x, y, z} represents the
anatomy
● Elements are called voxels, for volume
element, like picture element for pixels
● Standard anatomy has a resolution of
1mm
● At 7T, the resolution goes to ~0.3mm (27 times smaller,
because cubic relation between voxel size and volume!)
Other MRI data: What a versatile machine!
● Examples of 4D or higher-D images - store not a time-course for each voxel, but more
complex information
fMRI data: What does fMRI data look like?
2
,fMRI data: Data quality
● We are focused on 1. Signal, and 2. Noise - we want more signal, and less noise
● Measure for Scan Quality:
○ tSNR (temporal Signal to Noise Ratio)
μ
○ σ
, where μ is the mean image, and σ is the standard deviation over time
○ Normal values:
■ ~ 0, air, (μ close to 0)
■ > 40, noisy gray matter (more noise, higher σ)
■ > 80, high-signal gray matter (less noise, lower σ)
■ > 100, white matter (almost no noise, very low σ)
fMRI data: fMRI images as data
● Values in a normal fMRI dataset are not expressed in any SI unit, such as meter,
second, Tesla, etc.
● We cannot really compare across scanners, subjects, etc.
● Our workaround measure for BOLD signal strength:
○ Percentage signal change (%sc)
○ , where Ii is the image at timepoint i
○ Normal values:
■ 0.2%-1% (normal fMRI experiment @ 3T),
■ 1%-3% (finger-tapping in primary motor cortex @ 3T)
■ 1%-7% (finger-tapping in primary motor cortex @ 7T)
■ 7%-20% (voxel in a large vein in primary cortex @ 7T, where μ is very
low)
3
, fMRI analysis: How do you get to
results?
● Much more than just a single step
● Experimental design is
paramount:
● What is your goal?
● Many ways of going wrong
Replication Crisis: Many ways of going
wrong?
● With such complex analyses,
room for 1. ’Tinkering’, or 2.
Sloppiness
● We will teach you the newest approaches to reproducible neuroscience:
○ How to save your data
○ How to preprocess your data
○ How to program your data analyses
○ How to report your results
○ How to open-source your data and analyses
“Modern” fMRI: What’s the state of the art? - Naselaris et al, 2021 paper on Perusall
● With higher field strengths, more is possible. For instance, higher resolution means
“laminar” resolution
● Large, open science projects:
○ Human Connectome Project (HCP, 1200 sjs, >5 hours, 3T & 7T)
4
