15. Spatial and Temporal Resolution

Spatial and Temporal Resolution¶

Overview¶

Much of what can — and cannot — be inferred from an fMRI study traces back to its spatial and temporal resolution. Spatial resolution refers to the size of the voxels used to construct an image, and describes the ability to distinguish signal changes across anatomical locations. Temporal resolution describes the ability to distinguish signal changes across time points. Because data collection speed is limited, the two trade off against each other, and both are further constrained by the biophysics of the BOLD response itself.

Spatial resolution matters because neurons with similar functional properties tend to cluster anatomically, at scales ranging from whole-brain networks (centimeters) down to cortical columns (0.1–2 mm) and single cells (5–100 μm). Modern 3T studies typically acquire ~3 mm voxels, and high-field (7T) imaging can go below 1 mm — enough to reach a great deal of functionally meaningful structure. The thalamus, for example, contains over 30 nuclei per hemisphere averaging about 6.5 mm in diameter, comfortably within reach of standard fMRI. The finer the scale of interest, the more the analysis strategy matters: group analyses are sensitive to regions and networks, individual-subject analyses can localize signal down to 1–2 mm, and multivariate pattern analysis (MVPA) within individuals may pick up information encoded at even finer scales.

Given all that structure, why not simply acquire the smallest voxels the scanner allows? Three costs push back. First, the signal-to-noise ratio (SNR) of a voxel is proportional to its volume — halving each side of a 3 mm voxel cuts its volume, and thus its intrinsic SNR, by a factor of eight. Second, higher-resolution images take longer to acquire, degrading temporal resolution and increasing motion-related artifacts. Third, at the other extreme, voxels that are too large average over many neural populations and tissue types, diluting the signal from any one of them — the partial volume effect — and are more prone to susceptibility artifacts where they span tissue boundaries. The optimal voxel size balances these costs.

Even a perfectly chosen voxel size overstates what a study can resolve. The effective resolution is limited by the physiology of the BOLD response, which arises in arterioles and capillary beds and spreads into draining veins beyond the site of neural activity. This blurring is summarized by the BOLD point-spread function: roughly 3.5 mm at 3T and as low as ~1.5 mm at 7T, so there may be limited benefit in acquiring voxels much smaller than these values. Analysis choices erode resolution further, especially for group studies. Warping individual brains to a common template (e.g., MNI space) leaves substantial residual misalignment — functional areas like MT or the fusiform face area can vary in location by 2 cm or more across people — and spatial smoothing, applied partly to compensate, blurs the maps again. Reporting peak coordinates costs still more: across comparable group studies, peak locations for the same effect vary by roughly 2–3 cm. Approaches that localize regions or patterns of interest within individual participants, and functional alignment methods such as hyperalignment, preserve much more of the acquired resolution — including “hyperacuity” effects in which multivariate patterns capture information (e.g., ocular dominance) encoded at scales finer than the voxels themselves.

On the temporal side, fMRI is coarse compared with EEG or MEG. The sampling rate is set by the TR (typically 0.5–3 s), but the deeper limit is the hemodynamic response itself, which begins ~2 s after a brief event and peaks at 5–6 s. A TR of 2 s has therefore long been considered “adequate” — yet slow sampling carries real costs. The Nyquist theorem states that a periodic signal can only be resolved if it is sampled at more than twice its frequency; with sampling frequency $F_s = 1/\mathrm{TR}$, the highest resolvable frequency is

Any signal above this limit does not disappear — it is reflected (folded) back around the Nyquist frequency and masquerades as a lower-frequency signal, a phenomenon called aliasing. A 1 Hz heartbeat sampled at TR = 0.5 s ($f_{\mathrm{Nyquist}} = 1$ Hz) stays at its own frequency, well above the slow task-related band, where it can be filtered out. The same heartbeat sampled at TR = 2 s ($f_{\mathrm{Nyquist}} = 0.25$ Hz) is aliased into the low frequencies where task effects live, becoming inseparable from activation. Physiological noise is the largest noise source in fMRI and a major source of temporal autocorrelation, so aliased cardiac and respiratory signals are particularly damaging — for task analyses and even more so for functional connectivity. Slow sampling has a second cost: if event onsets are locked to the TR, the response peak can fall between samples. Presenting events at variable times relative to the TR oversamples the hemodynamic response and avoids this bias.

Designing an acquisition protocol means balancing four desirable properties — brain coverage, spatial resolution, temporal resolution, and freedom from artifacts — where improving one generally costs another. Whole-brain coverage at ~4 mm slice thickness takes about 2 s with standard protocols; better spatial or temporal resolution can be bought by reducing coverage, or by accelerated imaging: in-plane acceleration (skipping lines of k-space, with multi-coil reconstruction correcting the resulting aliasing artifacts) and multiband / simultaneous multi-slice (SMS) imaging, which excites and reads multiple slices at once. Multiband can reduce whole-brain TRs from ~2 s to a few hundred milliseconds — a current recommended standard is ~3 mm isotropic voxels, whole-brain, at TR < 1 s (e.g., multiband factor 6) — though each image is noisier, and aggressive acceleration can reduce contrast-to-noise and BOLD sensitivity. Many groups now push TR below 500 ms specifically to keep physiological signals below the Nyquist limit. Adopting a well-piloted standard protocol (e.g., harmonized with the Human Connectome Project or ABCD studies) reduces guesswork and aids comparability across studies.

Hands-on tutorial¶

The labs make these tradeoffs concrete with simulations. Here are two key steps: aliasing a “heartbeat” by sampling it at a typical TR, and smoothing small versus large activations.

Step 1 — Alias a heartbeat by sampling at the TR. We simulate a cardiac signal at 66 beats per minute (1.1 Hz) and sample it once per TR. At TR = 0.4 s the Nyquist limit ($1/(2 \cdot 0.4) = 1.25$ Hz) is above the cardiac frequency and the peak stays at 1.1 Hz; at TR = 2 s ($f_{\mathrm{Nyquist}} = 0.25$ Hz) it is aliased to 0.1 Hz — right into the task band.

% Adapted from CANlab tutorials (canlab.github.io) and
% github.com/canlab/FMRI_simulations
fs = 1000;                           % fs = simulation sampling rate (Hz) for "ground truth"
t = 0:1/fs:60-1/fs;                  % t = time vector, 60 s long
heart = sin(2*pi*1.1*t);             % simulated heartbeat: 66 bpm = 1.1 Hz

for TR = [0.4 2.0]                   % TR = time between volumes (s): fast vs. typical
    samp = heart(1:round(TR*fs):end);    % sample once per TR, like the scanner
    n = numel(samp);                     % n = number of samples (volumes)
    f = (0:floor(n/2)) ./ (n*TR);        % frequencies up to Nyquist
    p = abs(fft(samp)) / n;  p = p(1:floor(n/2)+1);
    [~, imax] = max(p);
    fprintf('TR = %.1f s: Nyquist = %.2f Hz, spectral peak at %.2f Hz\n', ...
        TR, 1/(2*TR), f(imax));
end

import numpy as np

fs = 1000                            # fs = simulation sampling rate (Hz) for "ground truth"
t = np.arange(0, 60, 1 / fs)         # t = time vector, 60 s long
heart = np.sin(2 * np.pi * 1.1 * t)  # simulated heartbeat: 66 bpm = 1.1 Hz

for TR in (0.4, 2.0):                # TR = time between volumes (s): fast vs. typical
    samp = heart[::round(TR * fs)]           # sample once per TR, like the scanner
    freqs = np.fft.rfftfreq(len(samp), d=TR) # frequencies up to Nyquist
    power = np.abs(np.fft.rfft(samp)) / len(samp)
    print(f"TR = {TR} s: Nyquist = {1/(2*TR):.2f} Hz, "
          f"spectral peak at {freqs[np.argmax(power)]:.2f} Hz")

Example output:

TR = 0.4 s: Nyquist = 1.25 Hz, spectral peak at 1.10 Hz
TR = 2.0 s: Nyquist = 0.25 Hz, spectral peak at 0.10 Hz

Step 2 — Smooth small vs. large activations. A 1-D “cortical strip” (1 voxel = 1 mm) contains two activations of equal amplitude: one narrow (~3.5 mm FWHM, the scale of a small nucleus or column cluster) and one broad (~19 mm). Smoothing with a typical 8 mm FWHM Gaussian kernel suppresses noise and preserves the broad activation, but flattens the narrow one — smoothing is a matched filter that favors signals at its own scale.

rng(1);                                    % seed for reproducible noise
x = (1:200)';                              % position along the strip: 1 voxel = 1 mm
small = 2*exp(-(x - 60).^2 / (2*1.5^2));   % narrow activation: ~3.5 mm FWHM, amplitude 2
large = 2*exp(-(x - 140).^2 / (2*8^2));    % broad activation: ~19 mm FWHM, amplitude 2
y = small + large + randn(200, 1);         % add Gaussian noise, sd = 1

fwhm = 8;                                  % fwhm = smoothing kernel FWHM (mm), a typical choice
sig = fwhm / (2*sqrt(2*log(2)));           % convert FWHM -> Gaussian sigma
kern = exp(-(-20:20).^2 / (2*sig^2))';  kern = kern / sum(kern);
ysmooth = conv(y, kern, 'same');

figure; plot(x, y, 'Color', [.7 .7 .7]); hold on
plot(x, ysmooth, 'k', 'LineWidth', 2)
plot(x, small + large, 'r--')
legend({'Noisy data' 'Smoothed (8 mm FWHM)' 'True signal'})
xlabel('Position (mm)')

from scipy.ndimage import gaussian_filter1d
import matplotlib.pyplot as plt

rng = np.random.default_rng(1)                 # seed for reproducible noise
x = np.arange(200)                             # position along the strip: 1 voxel = 1 mm
small = 2 * np.exp(-(x - 60)**2 / (2 * 1.5**2))   # narrow activation: ~3.5 mm FWHM, amplitude 2
large = 2 * np.exp(-(x - 140)**2 / (2 * 8**2))    # broad activation: ~19 mm FWHM, amplitude 2
y = small + large + rng.standard_normal(200)      # add Gaussian noise, sd = 1

fwhm = 8                                          # fwhm = smoothing kernel FWHM (mm), a typical choice
sigma = fwhm / (2 * np.sqrt(2 * np.log(2)))       # convert FWHM -> Gaussian sigma
y_smooth = gaussian_filter1d(y, sigma)

plt.plot(x, y, color=".7", label="Noisy data")
plt.plot(x, y_smooth, "k", lw=2, label="Smoothed (8 mm FWHM)")
plt.plot(x, small + large, "r--", label="True signal")
plt.xlabel("Position (mm)"); plt.legend()

Example output:

The full labs go further: building intuitions for sine waves and the FFT, aliasing a realistic spiky heartbeat with beat-to-beat variability, mapping which TRs keep cardiac signal separable from task frequencies, and quantifying how smoothing changes peak amplitude and SNR as a function of activation size.

Thought questions¶

1. You are planning two studies: one mapping responses across the ~30 nuclei of the thalamus (average diameter ~6.5 mm), the other decoding line orientation from columns in V1 (~1 mm scale). For each, choose a voxel size, field strength, coverage, and analysis strategy (group mapping, individual ROI localization, or MVPA), and justify each choice using SNR-per-voxel, partial volume effects, and the BOLD point-spread function.

2. A colleague argues: “The hemodynamic response peaks at 5–6 seconds, so there is no point sampling faster than TR = 2 s.” Drawing on aliasing, motion within volumes, HRF estimation, and per-image SNR, make the strongest case for and against a TR under 500 ms.

3. Effective resolution is set by the weakest link in a chain running from physiology (point-spread) through acquisition (voxel size) to analysis (normalization, smoothing, peak reporting, ROI averaging). For a standard group study — 3 mm voxels, MNI normalization, 8 mm smoothing, peak-coordinate tables — identify the weakest link, and describe which links you would change first to improve it, at what cost.

4. Smoothing acts as a matched filter: it boosts sensitivity to signals at its own spatial scale while attenuating finer ones. Given that the extent of true activations is usually unknown and varies across regions, how should a researcher choose a smoothing kernel — or should they smooth at all? Consider group alignment, partial volume effects, and the alternative of multivariate analyses on unsmoothed data.

5. MVPA can detect ocular dominance information using 2–3 mm voxels, even though the columns themselves are ~1 mm wide (“fMRI hyperacuity”). Explain how information at a scale finer than the voxel grid can survive sampling, and why group-average univariate maps of the same data would miss it.

Quiz yourself¶

⌂ Back to home · Table of contents · How to use this site