k-Space & the Fourier Transform

Why MRI does not measure pixels

A CT detector measures a quantity that maps almost directly onto a position in the patient. MRI does not. The receive coil integrates the entire transverse magnetization over the excited slice at every instant, returning a single complex voltage. What turns that bulk signal into an image is the way gradients tag spatial position as a phase that winds at a position-dependent rate. The data we actually record therefore live in a domain of spatial frequencies called k-space, and the image is recovered by an inverse Fourier transform. Understanding this single relationship explains field of view, resolution, aliasing, Gibbs ringing, partial-Fourier acceleration, and why the center of the data matters so much more than the edges.

This lesson builds the signal equation from the physics of a winding gradient phase, reads off the sampling rules that govern FOV and resolution, then uses k-space algebra to predict exactly what blurring, edge enhancement, image shifts, and aliasing look like before we ever reconstruct.

The signal equation: MRI samples a Fourier transform

Consider a slice with transverse spin density $m(x,y)$ (already weighted by T1, T2, and proton density). When a gradient $G_x(t)$ is applied, a spin at position $x$ precesses at an offset $\gamma G_x(t)\,x$ relative to isocenter, so by time $t$ it has accumulated an extra phase $\phi(x,t) = \gamma x \int_0^t G_x(\tau)\,d\tau$. Defining the running gradient area as a k-space coordinate collapses the messy time integral into a clean geometric variable.

The coil sums the contribution of every spin, each carrying phase $2\pi(k_x x + k_y y)$. The received signal is therefore the 2D Fourier transform of the spin density evaluated at the current $(k_x,k_y)$. As the gradients play out, the tip of the $(k_x,k_y)$ vector traces a path through k-space, and the ADC samples $S$ along that path.

The image is recovered by inverting Eq. 4.2. Reconstruction is, at heart, a 2D inverse fast Fourier transform of the sampled k-space matrix: $m(x,y) = \iint S(k_x,k_y)\, e^{+i2\pi(k_x x + k_y y)}\, dk_x\, dk_y$.

Center versus periphery: where contrast and detail live

Spatial frequency $|\mathbf{k}|$ is small near the center of k-space and large at the periphery. Low spatial frequencies describe slowly varying structure: bulk signal, large uniform regions, and overall image contrast. High spatial frequencies describe rapid variation: edges, fine texture, and sharp boundaries. Because the center sample $S(0,0) = \iint m\,dx\,dy$ equals the total integrated signal, it is by far the largest-magnitude point and dominates image brightness and contrast.

| k-space region | Spatial frequency | Encodes | Effect if removed | | --- | --- | --- | --- | | Center | Low | Bulk signal, contrast, coarse anatomy | Image loses nearly all signal and contrast | | Periphery | High | Edges, fine detail, sharp transitions | Image blurs but contrast is preserved |

Sampling rules: field of view and spatial resolution

k-space is sampled on a discrete grid with spacing $\Delta k$ out to a maximum extent $k_{max}$. Two Fourier-pair relationships, both consequences of the Nyquist sampling theorem applied in the k-domain, completely determine the imaging geometry. The sample spacing sets the field of view; the maximum extent sets the resolution.

The relationships are intuitive once the Fourier reciprocity is clear. Finer sampling (smaller $\Delta k$) extends the FOV, because closely spaced samples can resolve the slow phase ramps produced by objects far from isocenter without ambiguity. Reaching higher spatial frequencies (larger $k_{max}$) sharpens the image, because fine edges generate high-frequency content that only the outer reaches of k-space record. A worked example: to image a 256 mm FOV at 1 mm resolution along the readout, we need $\Delta k = 1/256\ \text{mm}^{-1}$ and $N = \mathrm{FOV}/\Delta x = 256$ samples, giving $k_{max} = N\,\Delta k / 2 = 0.5\ \text{mm}^{-1}$.

Reading k-space operations as image-domain effects

Because the two domains are an exact Fourier pair, a handful of Fourier theorems let us predict reconstruction outcomes without doing the transform by hand. These are the bread-and-butter mental models of MRI artifact analysis.

• Zero the periphery (low-pass filter): discarding high spatial frequencies convolves the image with a broad point spread function. The image blurs and edges soften, but overall contrast is preserved. This is the mechanism behind low-resolution scans and k-space apodization filters.
• Zero the center (high-pass filter): removing low frequencies leaves only edge information. The result is an edge-only, near-black image that resembles a gradient-magnitude map, with almost no bulk signal.
• Apply a linear phase ramp: multiplying k-space by $e^{-i2\pi k_x x_0}$ shifts the image by $x_0$ (the Fourier shift theorem). Gradient timing or eddy-current phase errors that look linear in k therefore appear as image displacement.
• Undersample (increase delta-k): spacing samples too far apart shrinks the effective FOV below the object, so signal wraps around as aliasing (the classic fold-over or wrap artifact in the phase-encode direction).
• Multiply k-space by a window (apodization): a smooth taper reduces ringing at the cost of a slightly wider PSF (mild blurring), trading sharpness for fewer truncation artifacts.

The interactive below lets you manipulate a real k-space matrix and watch the reconstruction update. Try cropping the periphery, masking the center, applying a phase ramp, and decimating the sampling to reproduce each effect above.

Hermitian symmetry and partial Fourier

If the image $m(x,y)$ were purely real, its Fourier transform would obey Hermitian (conjugate) symmetry: the sample at $-\mathbf{k}$ is the complex conjugate of the sample at $+\mathbf{k}$. In principle one half of k-space would then be redundant, and acquiring only one half plus a small overlap could halve the scan time. This is the basis of partial-Fourier (half-Fourier) acquisition.

In practice $m(x,y)$ is not purely real: B0 inhomogeneity, chemical shift, flow, and receive-coil phase impose a smooth spatial phase. The trick is that this phase varies slowly, so it can be estimated from a small symmetric band of fully sampled central lines. Sequences acquire roughly 55 to 75 percent of k-space (for example a partial-Fourier factor of 5/8), use the central lines to estimate and correct the background phase, then synthesize the missing half via Eq. 4.4. The cost is a $\sqrt{\text{fraction sampled}}$ reduction in SNR and some sensitivity to phase errors.

Trajectories, the point spread function, and Gibbs ringing

The path the $(k_x,k_y)$ vector traces is the trajectory, chosen by the gradient waveform. Cartesian sampling fills k-space line by line on a rectilinear grid and reconstructs by a direct FFT; it is robust and the workhorse of clinical imaging. Radial sampling acquires spokes through the center, oversampling the contrast-defining center on every readout, which makes it relatively motion-robust and tolerant of undersampling (streak rather than coherent wrap). Spiral trajectories start at the center and wind outward, covering k-space very efficiently per excitation, which makes them fast but sensitive to off-resonance blurring and gradient imperfections. Non-Cartesian data must be regridded onto a Cartesian matrix (or reconstructed with non-uniform FFT) before the standard inverse transform.

| Trajectory | Center sampling | Strengths | Main weakness | | --- | --- | --- | --- | | Cartesian | Once (one central line) | Simple FFT recon, predictable, robust | Coherent ghosting and wrap from motion or undersampling | | Radial | Every spoke | Motion-robust, benign undersampling (streaks) | Streak artifacts, needs regridding, more readouts | | Spiral | Start of each readout | Very fast, efficient k-space coverage | Off-resonance blurring, gradient-fidelity sensitive |

Whatever the trajectory, k-space is always finite: we stop at $k_{max}$. Truncating an infinite Fourier transform is equivalent to multiplying the ideal data by a rectangular window. By the convolution theorem, the reconstructed image is convolved with the Fourier transform of that window, the point spread function, which for a rectangle is a $\mathrm{sinc}$ function. The sinc has oscillating side lobes, so every sharp edge in the image rings: this is the Gibbs (truncation) artifact, seen as parallel bright and dark bands paralleling high-contrast boundaries.

Hold these four anchors and most of MRI data handling follows: the signal is a Fourier transform (Eq. 4.2); the center carries contrast and the periphery carries detail; FOV and resolution are set by $\Delta k$ and $k_{max}$ (Eq. 4.3); and every reconstruction artifact, from blur to aliasing to Gibbs ringing, is a predictable image-domain consequence of how k-space was sampled or modified.