fMRI
Functional magnetic resonance imaging. A method that uses repeated MRI images to study brain function indirectly through blood-related signal changes.
A practical neuroimaging course
Learn fMRI the way you would read a scan: signal first, map second, interpretation last.
Functional MRI, or fMRI, uses repeated MRI images to infer brain activity from blood oxygen changes. You will first learn the beginner vocabulary, including voxel, BOLD, HRF, GLM, contrast, threshold, and connectivity, then build toward preprocessing, statistical maps, open datasets, and advanced spatial-navigation signals.
Module 0
First, Learn The Language
fMRI becomes much less intimidating when every technical word has a plain meaning. Keep this page open as your decoder. The goal is not to memorize jargon. The goal is to know what question each word helps you ask.
The whole fMRI story in five sentences
- A scanner takes repeated 3-D snapshots of the brain.
- Each snapshot is made of many voxels, which are tiny 3-D measurement boxes.
- Each voxel has a time series, meaning its signal value across many snapshots.
- Neural activity changes local blood oxygen, producing the delayed BOLD signal.
- Analysis asks whether the signal pattern matches a task, a network, or a prediction.
Beginner rule: when you see a colored brain image, ask three questions:
what was measured, what was done to the data, and what comparison made the color appear?
BOLD
Blood-oxygen-level dependent signal. It changes because oxygenated and deoxygenated blood affect the MRI signal differently.
Voxel
A volume pixel, or a tiny 3-D cube of brain image. A voxel contains many neurons, blood vessels, and other tissue, not one neuron.
Volume
One 3-D snapshot of the brain made from many slices and voxels. A run contains many volumes collected over time.
Time series
The list of signal values from the same voxel or region across the scan. Most fMRI analysis starts with time series.
TR
Repetition time. The time between one brain volume and the next. If TR is 2 seconds, the scanner samples the brain every 2 seconds.
HRF
Hemodynamic response function. The expected delayed BOLD shape after neural activity: slow rise, peak, and return toward baseline.
Preprocessing
The cleanup and alignment steps before statistics: motion correction, distortion correction, registration, normalization, smoothing, and denoising.
GLM
General linear model. A regression model that asks whether a voxel time series follows the timing of a task or condition.
Design matrix
The table of predictors used by the GLM. Each column represents something that may explain the signal, such as a task condition or motion.
Contrast
A specific comparison, such as faces greater than houses. It turns model estimates into an interpretable question.
Threshold
A cutoff used to decide which statistical values are displayed or treated as evidence. Changing it can change the visible map.
Correction
A statistical safeguard for testing thousands of voxels. It reduces false positives caused by asking many questions at once.
ROI
Region of interest. A chosen brain area used to summarize signal or test a focused hypothesis.
Connectivity
A measure of how similarly two time series change. Functional connectivity does not prove direct wiring or causation.
Confound
Something unwanted that can affect the signal, such as head motion, breathing, heartbeat, scanner drift, or site differences.
Beginner sentence
"The red blob means this brain area lit up."
Analyst sentence
"This corrected contrast showed higher BOLD signal in this cluster for condition A compared with condition B."
Beginner sentence
"These regions are connected, so one causes the other."
Analyst sentence
"These regions had correlated BOLD time series in this sample, which suggests functional coupling but not direct causality."
Module 1
What fMRI Measures
Functional MRI does not measure neurons directly. It measures an MRI signal that changes with blood oxygenation, blood volume, and blood flow after neural activity. The most common signal is BOLD: blood-oxygen-level dependent contrast.
The core chain
- Neurons consume energy during computation.
- Local vessels over-deliver oxygenated blood after a short delay.
- Deoxyhemoglobin changes local magnetic properties.
- The scanner detects a tiny signal change, often around 0.5 to 5 percent.
- Analysis converts repeated images into time series, maps, and statistics.
Interpretation rule: fMRI is best read as an indirect, delayed, spatially
local but physiologically mixed signal. Never say "this voxel fired" when you mean
"this voxel showed a BOLD contrast."
Module 2
From Anatomy To Voxels
An fMRI run is a stack of 3-D brain volumes sampled repeatedly over time. Each tiny cube is a voxel, which is easiest to imagine as a 3-D pixel. For every voxel, the scanner records a time series. Interpretation starts by knowing whether you are looking at anatomy, a single functional volume, an averaged activation map, or a statistical contrast.
Anatomical MRI
High-resolution structure. Useful for localization, segmentation, normalization, and checking whether the brain looks anatomically plausible.
Functional run
Many fast, lower-resolution volumes. Useful for time-series analysis, activation, and connectivity.
Contrast map
A derived image showing a comparison between conditions or model estimates. This is what most published activation figures show.
Slice lab
Move through a teaching brain volume and adjust the overlay threshold. This is simulated for learning, but it mirrors the questions you ask when reading real maps.
Module 3
BOLD, HRF, And Time
The BOLD response is delayed and blurred in time. A neural event can happen in milliseconds, but the hemodynamic response usually rises after a few seconds, peaks later, and returns slowly. That delay is why fMRI is strong for spatial maps but weak for precise timing.
Plain translation: the brain event is quick, but the blood response is slow.
fMRI is like watching the wake after a boat has passed, not the boat itself.
Terms you must know
- TR
- Repetition time. How often a whole-brain volume is sampled, usually measured in seconds.
- TE
- Echo time. The time between excitation and signal readout. It affects BOLD sensitivity and image contrast.
- HRF
- Hemodynamic response function. The expected BOLD shape after neural activity.
- Percent signal change
- A normalized way to talk about small BOLD changes.
- Drift
- Slow signal change over time that can come from scanner or physiology rather than the task.
- Motion spike
- A sudden signal jump caused by head movement. It can imitate brain effects if ignored.
Interpretation rule: a BOLD peak does not mark the moment neurons became active.
It is a delayed vascular signature related to earlier neural activity.
Module 4
How fMRI Data Are Organized
Most modern public neuroimaging datasets use BIDS, the Brain Imaging Data Structure. BIDS gives you predictable folders, filenames, JSON metadata, event timing files, and derivative folders. If you can read BIDS, you can approach many open fMRI datasets without getting lost.
dataset_description.json
participants.tsv
sub-01/anat/sub-01_T1w.nii.gz
sub-01/func/sub-01_task-memory_bold.nii.gz
sub-01/func/sub-01_task-memory_events.tsv
derivatives/fmriprep/sub-01/func/...preproc_bold.nii.gz
NIfTI
The image file format. A 4-D BOLD file stores left-right, front-back, up-down, and time.
JSON sidecar
A small metadata file beside the image. It can tell you TR, task name, slice timing, phase encoding, and scanner settings.
Events TSV
A table of when task events happened. This is what lets the GLM connect the task design to the BOLD time series.
Open-data atlas
Datasets Worth Learning From
Use these resources to move from course diagrams to real data. Some are fully open, while others require a data-use agreement. Start with OpenNeuro or NeuroVault if you want the lowest barrier to practice.
OpenNeuro
BIDS-formatted public neuroimaging datasets. Best for learning BIDS, task fMRI, resting-state fMRI, and reproducible workflows.
Open repositoryNeuroVault
Public statistical maps, atlases, parcellations, and unthresholded MRI/PET-derived maps. Best for learning map interpretation.
Open mapsHuman Connectome Project
High-quality structural, task, resting-state, diffusion, and MEG data for healthy adults. Best for connectivity and networks.
View HCPNatural Scenes Dataset
Large 7T fMRI dataset from subjects viewing thousands of natural scenes. Best for visual cortex, encoding models, and decoding.
View NSDStudyForrest
Naturalistic fMRI around movie watching and audio narratives. Best for time-locked natural stimulus analysis.
View StudyForrestABIDE
Autism Brain Imaging Data Exchange. Best for multi-site resting-state fMRI, site effects, motion, and clinical interpretation caution.
View ABIDE
Dataset rule: before analysis, read the dataset paper, task description,
participant table, exclusion criteria, scanner parameters, and preprocessing status.
Module 5
Preprocessing: Making Images Comparable
Preprocessing reduces nuisance variation and puts images into a form that can be modeled. It does not magically remove all artifacts. Every preprocessing choice can change your interpretation.
Plain translation: preprocessing is the preparation stage. It tries to make
images line up, reduce known noise, and produce data that can be compared across time,
people, and brain regions.
1Discard unstable volumes
2Slice timing correction
3Motion correction
4Distortion correction
5Coregister to anatomy
6Normalize to template
7Spatial smoothing
8Denoise and model
Discard unstable volumes
Remove the first few volumes because the scanner signal may not be steady yet.
Slice timing correction
Adjust for the fact that slices in one volume were not all captured at the exact same moment.
Motion correction
Estimate and correct small head movements so the same voxel refers to the same brain location over time.
Coregistration
Align functional images to the participant's anatomical image so statistics can be localized.
Normalization
Warp each participant into a common template space so group analysis becomes possible.
Smoothing
Average nearby voxels slightly to improve signal-to-noise, while accepting some loss of spatial sharpness.
Motion
Head movement can create false activation and false connectivity. Check framewise displacement and motion outliers.
Physiology
Breathing and heartbeat affect BOLD. They can mimic networks if ignored.
Susceptibility
Orbitofrontal and temporal regions are vulnerable to signal dropout and distortion.
Registration
Bad alignment makes anatomical interpretation untrustworthy, even if the statistics look clean.
Module 6
Task fMRI And The GLM
The general linear model asks whether a voxel time series is better explained by task predictors than by noise. A contrast then asks a sharper question, such as faces greater than houses, memory encoding greater than baseline, or left button press greater than right.
Write the question
Example: does navigation produce more BOLD signal than a control condition?
Build predictors
Turn event timing into model columns, then add nuisance columns such as motion.
Estimate betas
For each voxel, estimate how strongly each predictor explains the time series.
Test a contrast
Compare beta weights to create a statistical map, such as navigation greater than control.
Design matrix
Contrast preview
Beta
Estimated contribution of one predictor to the voxel signal. Larger beta means the model thinks that predictor explains more of the signal.
Contrast
A weighted question about betas. Example: navigation greater than control, or faces greater than houses.
Statistical map
A voxelwise map of evidence for the contrast, often shown as t, z, or F values.
Module 7
Reading Activation Maps
Activation maps are not photographs of thought. They are statistical answers to a model. You need the contrast, threshold, correction method, sample size, and anatomical localization before you interpret a colored blob.
What contrast is shown?
A map without a contrast is uninterpretable.
What threshold was used?
Uncorrected maps can look persuasive while being false positive heavy.
Where is the cluster?
Check atlas labels, coordinates, and individual anatomy.
Could it be artifact?
Motion, vessels, dropout, and smoothing can fool interpretation.
Threshold
The cutoff for showing or trusting a statistical value. A lower threshold shows more, but it can also show more false positives.
Multiple comparisons
fMRI tests thousands of voxels. Correction methods reduce the chance that random noise looks meaningful somewhere.
Cluster
A neighboring group of voxels that pass the threshold. A cluster is not automatically a single brain area.
Atlas label
A named anatomical or functional label used to describe where a cluster is. Always treat it as a guide, not a guarantee.
Interpretation rule: never interpret "red" as "important." Interpret a
corrected statistical contrast in anatomical and experimental context.
Module 8
Resting-State fMRI And Connectivity
Resting-state fMRI studies spontaneous low-frequency BOLD fluctuations. Regions whose time series fluctuate together are said to be functionally connected. This does not prove direct anatomical wiring or causality.
Seed-based
Pick one region of interest and correlate its time series with the rest of the brain.
ICA
Independent component analysis. A data-driven method that separates the scan into spatial networks and noise-like components.
Graph theory
Represent brain regions as nodes and connectivity values as edges, then study the network's organization.
Module 9
Common Mistakes In fMRI Interpretation
Mistake
Better interpretation
"The amygdala lit up."
"A contrast showed higher BOLD signal in an amygdala-region cluster under this model."
"This network causes memory."
"Connectivity in this network is associated with memory performance in this sample."
"No activation means no neural activity."
"No significant BOLD contrast was detected under this design and threshold."
"Group maps describe every participant."
"Group maps summarize a population effect and can hide individual variability."
"A voxel is a neuron."
"A voxel is a 3-D measurement unit that contains many cells, vessels, and possible sources of noise."
"Functional connectivity means direct wiring."
"Functional connectivity means time series similarity. It can reflect indirect paths, shared inputs, physiology, or artifacts."
Module 10
From Maps To Biomarkers
A biomarker is not just a significant group difference. It should be reliable, interpretable, testable in new data, and useful for prediction, diagnosis, prognosis, or mechanism. fMRI biomarkers face major challenges: motion, scanner differences, small samples, analysis flexibility, and weak external validation.
Biomarker
A measurable signal that helps classify, predict, monitor, or explain a biological or clinical state.
Reliability
Whether the measure is stable enough to repeat within the same person, scanner, or study design.
Validity
Whether the measure actually captures the process it claims to capture.
External validation
Testing the marker in a new dataset or site, not only in the dataset used to discover it.
ReliabilityDoes it repeat within subject and across sites?
ValidityDoes it measure the process you claim?
PredictionDoes it work in unseen participants?
Clinical valueDoes it improve decisions beyond cheaper measures?
Advanced module
Grid-Like Signals And Spatial Navigation fMRI
Grid cells in entorhinal cortex fire in a hexagonal spatial pattern in animals. Human fMRI cannot directly see individual grid cells, but some navigation tasks estimate a grid-like population signal: BOLD activity varies with movement direction in a six-fold pattern.
Hexadirectional signal lab
Rotate the virtual path. A simplified grid-like model predicts stronger signal when movement aligns with a six-fold orientation. Real analyses require careful controls for speed, visual input, task timing, smoothing, and circular statistics.
How to read grid-like fMRI claims
- Check whether the task truly requires navigation or path integration.
- Check how movement direction was modeled.
- Check whether the six-fold model was compared with control symmetries.
- Check whether the effect is localized to entorhinal or medial temporal regions.
- Check whether the result survives motion, speed, and visual confounds.
Advanced interpretation rule: a grid-like fMRI signal is a population-level
directional modulation, not proof that individual human grid cells were recorded.
Reference module
fMRI Glossary You Can Return To
Use this as a quick reference while reading papers, browsing datasets, or interpreting maps. The expert move is to define the term, state what it allows you to infer, and state what it does not prove.
- Activation
- A statistically modeled BOLD difference for a task or condition. It is not direct neural firing.
- Atlas
- A labeled brain reference used to name regions or networks.
- Beta
- The GLM estimate for how strongly one predictor explains a voxel's time series.
- BOLD
- Blood-oxygen-level dependent signal, an indirect MRI signal linked to blood oxygenation changes.
- Confound
- An unwanted influence such as motion, breathing, heartbeat, scanner drift, or site effects.
- Contrast
- A planned comparison of model estimates, such as task greater than control.
- Design matrix
- The GLM table whose columns represent task predictors and nuisance predictors.
- FWE / FDR
- Common correction approaches for limiting false positives across many voxelwise tests.
- GLM
- General linear model, the regression framework commonly used for task-fMRI analysis.
- HRF
- Hemodynamic response function, the delayed BOLD response shape expected after neural activity.
- ICA
- Independent component analysis, a data-driven method that separates signal into components.
- NIfTI
- The common neuroimaging file format for structural and functional MRI images.
- ROI
- Region of interest, a selected brain area used for focused measurement or hypothesis testing.
- Seed
- A chosen region whose time series is compared with other regions in seed-based connectivity.
- Statistical map
- A brain image where each voxel stores a statistic, often from a GLM contrast.
- Threshold
- A cutoff for display or inference. Always report it because it affects what appears on the map.
- Time series
- A sequence of signal values measured from the same voxel or region over time.
- Voxel
- A small 3-D image unit. It contains many biological and measurement sources, not one cell.
Fluency test: for any fMRI claim, you should be able to say the data type,
the preprocessing status, the model or connectivity method, the threshold or correction,
and the cautious interpretation.
Capstone
Your fMRI Interpretation Checklist
Where to go next
Start with a small OpenNeuro task dataset, inspect its BIDS files, open a contrast map in NeuroVault, then read a paper that uses HCP or NSD. Once the basics feel natural, return to spatial navigation and grid-like signals as an advanced special topic.