Functional magnetic resonance imaging Study Guide

📖 Core Concepts Functional MRI (fMRI) – non‑invasive imaging that maps brain activity by detecting changes in blood flow (neurovascular coupling) using the BOLD (blood‑oxygen‑level‑dependent) contrast. Neurovascular coupling – neuronal firing → glutamate release → astrocytic Ca²⁺ rise → nitric‑oxide‑mediated arteriolar dilation → influx of oxygenated blood, lowering deoxy‑hemoglobin (dHb). BOLD signal – T2 signal increase when dHb concentration falls; dHb is paramagnetic (causes signal loss), oxygenated Hb is diamagnetic (minimal loss). Hemodynamic Response Function (HRF) – stereotyped BOLD timecourse: onset ≈ 2 s, peak 4‑6 s, possible undershoot after stimulus offset; modeled as a canonical impulse response. Linear superposition – For modest stimulus spacing, the total BOLD response ≈ sum of individual HRFs (convolution of stimulus sequence with HRF). General Linear Model (GLM) – statistical framework that fits the predicted BOLD timecourse (design matrix × HRF) to each voxel’s observed signal to estimate task‑related activation. 📌 Must Remember BOLD ∝ (B₀)² – effect grows with the square of the static magnetic field strength (≥ 1.5 T). HRF timing: onset ≈ 2 s, peak ≈ 5 s, undershoot ≈ 10‑12 s. Spatial resolution limited by voxel size (≈ 4‑5 mm³ for whole‑brain, sub‑mm for laminar) and large draining veins. Temporal resolution limited by repetition time (TR); decreasing TR < 1 s gives little extra temporal precision because the HRF is >10 s wide. Linear range: up to 2 s inter‑stimulus interval; beyond this refractory period causes non‑linearity. Noise scaling: physiological noise ∝ (B₀)²; thermal noise is constant across field strengths. Pre‑processing order: slice‑time → motion correction → distortion correction → coregistration → normalization → spatial smoothing → temporal filtering. Statistical thresholding must control family‑wise error or false‑discovery rate; uncorrected voxel‑wise p < 0.001 is insufficient for whole‑brain inference. 🔄 Key Processes Data Acquisition Gradient‑echo EPI → T2‑weighted volumes every TR. Pre‑processing Slice timing correction → align slice acquisition times. Rigid‑body motion correction → estimate translation/rotation to first volume. Distortion correction (field map) → remove B₀ inhomogeneities. Coregister functional to high‑resolution structural image. Normalize to standard space (MNI/Talairach). Spatial smoothing (Gaussian kernel ≈ FWHM 4‑8 mm). Temporal filtering (high‑pass ≥ 1/(2 TR), optional low‑pass). Design Matrix Construction Encode each experimental condition as a binary time series. Convolve each column with canonical HRF → predicted BOLD regressors. GLM Fitting Solve β = (XᵀX)⁻¹ Xᵀ y (least‑squares) for each voxel, where X is the design matrix, y the observed timecourse. Statistical Inference Compute t‑ or F‑statistics for contrasts of β‑weights. Apply multiple‑comparison correction (e.g., cluster‑wise FWE, TFCE). Interpretation Thresholded activation maps → brain regions linked to task. Optional MVPA: train classifier on voxel patterns, test on independent runs. 🔍 Key Comparisons Oxygenated Hb (HbO₂) vs Deoxygenated Hb (dHb) – diamagnetic vs paramagnetic → opposite effects on T2 signal. Gradient‑echo EPI vs Spin‑echo EPI – gradient‑echo: high BOLD sensitivity, includes large‑vein signal; spin‑echo: suppresses large veins, improves spatial specificity. Block design vs Event‑related design – block: higher power, vulnerable to drift; event‑related: captures transient responses, lower per‑event SNR. Spatial smoothing vs No smoothing – smoothing ↑ SNR & meets Gaussian random field assumptions; excessive smoothing blurs fine‑scale activations. Linear GLM vs Non‑linear models – GLM assumes additivity & stationarity; non‑linear models needed for short ISIs (< 2 s) or saturation. ⚠️ Common Misunderstandings “BOLD = neuronal firing.” – BOLD correlates more strongly with local field potentials (synaptic input) than with spiking output. “Resting‑state = no brain activity.” – Resting‑state BOLD shows organized networks (default mode). “Higher BOLD = better performance.” – Efficiency can reduce BOLD amplitude while performance stays constant. “fMRI can pinpoint the exact neuron.” – Spatial specificity limited by voxel size and venous contributions; cannot resolve single‑cell activity. “One subject’s map is diagnostic.” – fMRI is reliable for group‑level inference; individual‑level reliability is poor. 🧠 Mental Models / Intuition “BOLD as a flood detector.” – Neuronal activation opens a “gate” (vasodilation) that floods the area with oxygenated blood, washing out the “mud” (dHb) that darkens the MR image. “HRF as a camera shutter.” – The brain’s vascular response is a delayed, blurry exposure of the underlying neural event; convolving the stimulus with the shutter (HRF) gives the picture you record. “Linear superposition = stacking transparent slides.” – Each stimulus adds a semi‑transparent HRF slide; the final image is the sum of all slides unless they overlap too much (refractory period). 🚩 Exceptions & Edge Cases Short inter‑stimulus intervals (< 2 s): refractory period reduces BOLD amplitude → linearity breaks. Very high field (> 3 T): physiological noise dominates, diminishing returns on SNR. Pathology (tumors, stroke): altered vascular reactivity can decouple BOLD from neural activity. Drugs, anxiety, sedation: modify neurovascular coupling → BOLD amplitude/shape changes. Large draining veins: contribute BOLD signal far from neuronal source; spin‑echo or high‑field gradient‑echo can mitigate. 📍 When to Use Which Gradient‑echo EPI → standard whole‑brain BOLD studies where sensitivity > spatial specificity. Spin‑echo EPI → high‑resolution laminar or cortical column work; need to suppress vein signal. Block design → hypothesis testing with strong expected effects; limited number of conditions. Event‑related design → trial‑by‑trial analyses, deconvolution, or when stimulus order matters. Spatial smoothing (4‑6 mm FWHM) → typical group analyses; use smaller kernels (< 3 mm) for high‑resolution or MVPA. High‑pass filter cutoff = 1/(2 TR) → remove slow drifts; adjust upward if block design introduces low‑frequency variance. 👀 Patterns to Recognize Canonical HRF shape appearing repeatedly across voxels → true task‑related activation. Clustered activation in large veins (linear, extended along sagittal sinus) → vascular artifact, not focal neural signal. Motion‑related spikes coinciding with sudden signal jumps → need motion regressors or volume censoring. Physiological noise peaks at cardiac (1 Hz) and respiratory (0.3 Hz) frequencies → appropriate band‑pass filtering can improve SNR. Consistent activation across subjects in the same anatomical location → robust group effect; divergent patterns may indicate subject‑specific variability or noise. 🗂️ Exam Traps “BOLD signal directly measures neuronal firing.” – distractor; correct answer emphasizes synaptic input/LFP correlation. “Higher field always yields better data.” – trap; ignore increased physiological noise at > 3 T. “TR = 0.5 s will double temporal precision.” – false; HRF limits effective temporal resolution. “Block designs are always superior.” – mislead; block designs are vulnerable to drift and cannot separate transient processes. “Spatial smoothing always improves detection.” – over‑smoothing can merge adjacent activations and violate assumptions; optimal kernel matches expected activation size. “A significant voxel after uncorrected p < 0.001 proves activation.” – ignore multiple‑comparison correction; cluster‑level FWE needed. --- Use this guide to quickly recall the most exam‑relevant facts, workflows, and pitfalls of functional MRI.