Signal Localisation and Image Formation (k-Space)

Version 1.0 — July 2026

1. Executive Summary

This page covers, in depth, the third of the eleven planned topic groups listed in Section 4.3 of the MRI Physics — Fundamentals and Principles master page: how a raw MRI signal, which by itself carries no spatial information at all, is converted into a spatially resolved image. Five tightly linked topics are covered, following the master page’s own roadmap for this group: the three magnetic field gradients that provide spatial encoding, the physics of slice selection, the theory and structure of k-space, the major strategies used to fill k-space (Cartesian, non-Cartesian, and partial Fourier), and the Fourier transform that ultimately converts the acquired k-space data into a viewable image — together with the physical basis of the fundamental resolution/SNR/scan-time trade-off that governs essentially every protocol decision documented elsewhere on MRIninja.

2. Relationship to the Master Page

The master page introduces spatial encoding and k-space only briefly (Section 3.4), and the companion child pages Fundamentals of Nuclear Magnetic Resonance (which establishes that a raw FID signal carries no spatial information whatsoever) and RF Pulses and Pulse Sequence Physics (which introduces frequency-selective RF pulses) are both assumed as background here. This page adds: the detailed physics of how three orthogonal gradients (slice-select, phase-encode, frequency-encode) jointly localise signal in three dimensions, the formal mathematical relationship between applied gradients and k-space coordinates, the practical trade-offs of different k-space sampling strategies, and the physical (not merely mathematical) basis of the resolution/SNR/scan-time relationship that underlies protocol optimisation throughout MRIninja’s anatomical and parameter pages.

3. Magnetic Field Gradients — The Three Encoding Axes

3.1 The Common Principle — Position-Dependent Larmor Frequency

As established in the companion Fundamentals of Nuclear Magnetic Resonance child page, the Larmor equation makes precession frequency directly proportional to local magnetic field strength. A gradient coil deliberately superimposes a small, spatially linear variation in field strength on top of the uniform B0 field along a chosen axis, so that the local Larmor frequency (and, during the gradient’s application, the accumulated local phase) becomes a linear, predictable function of position along that axis. All three spatial encoding mechanisms described below are direct applications of this single physical principle to a different stage of the pulse sequence timeline and a different physical observable (frequency vs phase).

3.2 Slice-Select Gradient

Applied simultaneously with a frequency-selective RF excitation pulse (introduced in the companion RF Pulses and Pulse Sequence Physics child page), the slice-select gradient ensures that only the thin slab of tissue whose local Larmor frequency falls within the RF pulse’s bandwidth is excited — the direct physical mechanism of slice selection, detailed further in Section 4.

3.3 Frequency-Encode (Readout) Gradient

Applied during actual signal reception (not during excitation), the frequency-encode gradient — also called the readout gradient — makes the precessing signal’s instantaneous frequency a linear function of position along its axis. Because the receiver samples this composite signal as a function of time, and different spatial positions along the readout axis are, by construction, oscillating at different frequencies during acquisition, subsequent Fourier analysis of the sampled time-domain signal (Section 6) directly separates contributions from different positions along this axis.

3.4 Phase-Encode Gradient

Applied briefly, before signal readout, and (critically) with a different amplitude on each repetition of the pulse sequence, the phase-encode gradient imprints a position-dependent phase shift — rather than a position-dependent frequency shift — on the transverse magnetization along its axis. Because a single measurement can only capture the phase state that exists at one moment, distinguishing different positions along the phase-encode axis requires repeating the entire excitation-and-readout process multiple times, each time with a different phase-encode gradient amplitude — the direct physical reason phase-encoding, rather than frequency-encoding, is the primary determinant of total scan time in conventional Cartesian 2D and 3D imaging (Section 6.1), and the reason reducing the number of phase-encode steps (through undersampling, partial Fourier, or parallel imaging) is such a central strategy for accelerating MRI acquisition.

4. The Physics of Slice Selection

Slice selection combines a frequency-selective RF pulse (Section 4.2 of the companion RF Pulses and Pulse Sequence Physics child page) with a simultaneously applied slice-select gradient (Section 3.2 above). The excited slice’s thickness is determined jointly by the RF pulse’s bandwidth and the slice-select gradient’s strength: a stronger gradient spreads a given range of Larmor frequencies over a shorter physical distance, producing a thinner slice for the same RF pulse bandwidth, while a narrower-bandwidth RF pulse produces a thinner slice for the same gradient strength. In practice, achieving a genuinely rectangular slice profile (uniform excitation across the intended slice thickness, with a sharp cutoff at the edges) is limited by the same time-bandwidth trade-off discussed in the companion RF Pulses page (Section 10.1 there): real, finite-duration RF pulses always produce some degree of imperfect slice profile, most pronounced at the slice edges, with practical consequences (partial-volume-like signal blending between adjacent slices) that are managed clinically through interslice gap conventions and pulse-shape optimisation rather than eliminated outright.

5. k-Space — Theory and Structure

5.1 What k-Space Represents

k-space is the raw data matrix actually acquired during an MRI scan — not an image, but the two- (or three-) dimensional spatial-frequency domain representation of the eventual image, related to it by the Fourier transform (Section 6). Each point in k-space does not correspond to a single point or region in the final image; rather, each acquired k-space data point contains information contributed by every point in the imaged object simultaneously, weighted according to that k-space location’s specific spatial frequency. This is the fundamental reason a single corrupted or missing k-space data point produces a subtle, diffuse effect across the entire reconstructed image (such as faint streaking) rather than a localised defect at one image position — the exact opposite of how errors behave in the final, real-space image itself.

5.2 The Gradient–k-Space Relationship

The formal mathematical relationship connecting applied gradients to k-space position, k(t) = γ∫G(t’)dt’ (integrating the applied gradient waveform over time, scaled by the gyromagnetic ratio), was formalised independently by Ljunggren and by Twieg in 1983, building on the spin-warp imaging framework Edelstein and colleagues had described three years earlier [1,2,3]. This relationship is the precise, quantitative link between the physical hardware event (a gradient waveform, in units of field-strength-per-distance, applied for a certain duration) and the resulting position in the abstract k-space data matrix that is actually being filled in at that moment — every distinct pulse sequence and k-space trajectory (Section 6) is, at its core, simply a different deliberate choice of gradient waveforms designed to trace out a particular desired path through k-space over time.

5.3 Structural Properties of k-Space

Two structural properties of k-space have direct, practical consequences discussed throughout MRIninja’s protocol pages:

• Centre vs periphery: k-space data near the centre (low spatial frequency) predominantly determines overall image contrast and signal-to-noise, while data toward the periphery (high spatial frequency) predominantly determines fine spatial detail and edge sharpness. This is the physical basis for k-space-weighted acquisition strategies (such as centric or elliptical-centric ordering) that deliberately acquire the contrast-determining central region at a particular point in the contrast-dynamics timeline.
• Conjugate (Hermitian) symmetry: for an object with no phase errors, k-space possesses a mathematical symmetry in which data at one location and data at the mirror-image location across the k-space centre are related (specifically, complex conjugates of one another) — meaning, in principle, only slightly more than half of k-space needs to be directly measured, with the remainder inferable by symmetry. In practice, genuine phase errors (from field inhomogeneity, motion, flow, and coil-related effects, among other sources) mean this symmetry is never perfectly exact, which is the direct physical reason partial Fourier techniques (Section 6.3) acquire somewhat more than the theoretical minimum half of k-space.

6. k-Space Filling Strategies and Trajectories

6.1 Cartesian Sampling

The great majority of clinical MRI sequences fill k-space on a regular Cartesian grid, one line at a time, with each line’s position along the phase-encode direction set by that repetition’s phase-encode gradient amplitude (Section 3.4) and its extent along the frequency-encode direction filled continuously during readout (Section 3.3) — the direct descendant of the spin-warp method described by Edelstein and colleagues in 1980. Cartesian sampling’s principal practical advantage is straightforward, computationally efficient reconstruction via the standard fast Fourier transform (Section 6); its principal limitation is that total scan time scales directly and linearly with the number of phase-encode lines acquired.

6.2 Non-Cartesian Trajectories — Radial and Spiral

Non-Cartesian k-space trajectories acquire data along paths other than a rectangular grid. Radial trajectories, conceptually related to Lauterbur’s original 1973 projection-reconstruction proposal, acquire k-space data along lines passing through (or near) the k-space centre at varying angles, analogous to the projection geometry of CT; a practical advantage is that the densely, repeatedly sampled k-space centre confers relative motion robustness compared with Cartesian sampling, at the cost of a more computationally demanding (typically gridding-based) reconstruction. Spiral trajectories acquire k-space data along outward- (or inward-) spiralling curved paths from the centre, offering efficient k-space coverage per unit acquisition time, at the cost of increased sensitivity to off-resonance and gradient-timing-related image blurring artefact relative to Cartesian sampling. The general mathematical framework describing how the detailed time-course of any such trajectory determines image quality and artefact behaviour — applicable equally to Cartesian, radial, spiral, and echo-planar sampling — was formalised in Twieg’s 1983 k-trajectory paper [2].

6.3 Partial Fourier (Half-Fourier) Acquisition

Exploiting the conjugate symmetry property described in Section 5.3, partial Fourier (also called half-Fourier) acquisition deliberately samples only slightly more than half of k-space along a given encoding direction, mathematically estimating the remaining, unacquired portion using the measured symmetric data — a technique independently described by Feinberg and colleagues and by Margosian in 1985–1986 [4]. Because real phase errors mean the symmetry is never perfectly exact (Section 5.3), practical partial Fourier implementations acquire a modest additional margin beyond the theoretical minimum (commonly around five-eighths to seven-eighths of full k-space, rather than exactly one-half) specifically to allow estimation and correction of these phase errors from the acquired data itself, at some cost to the theoretical maximum time saving.

7. The Fourier Transform, Image Reconstruction, and the Resolution/SNR/Scan-Time Relationship

7.1 From k-Space to Image

Because k-space is, by construction, the spatial-frequency-domain representation of the object being imaged, the final image is obtained by applying an inverse Fourier transform to the acquired (and, where relevant, partial-Fourier-completed or otherwise processed) k-space data — the mathematical relationship first proposed as a practical NMR imaging strategy in Lauterbur’s foundational 1973 paper [5]. For standard Cartesian data, this is computed efficiently using the fast Fourier transform algorithm; non-Cartesian trajectories (Section 6.2) typically require an additional resampling (“gridding”) step onto a regular grid before the same fast Fourier transform machinery can be applied.

7.2 The Resolution/SNR/Scan-Time Relationship — Physical Basis

Three of the most fundamental, universally applicable trade-offs in MRI protocol design are direct physical consequences of the k-space framework described above, not arbitrary engineering conventions:

• Spatial resolution is set by k-space coverage. Because high spatial frequencies (fine image detail) correspond to the periphery of k-space (Section 5.3), achieving finer spatial resolution requires sampling further out into k-space — directly requiring either stronger gradients, longer/more numerous phase-encode steps, or both.
• Signal-to-noise ratio depends on the total amount of signal sampled. Each individual k-space data point is acquired with a certain amount of thermal noise superimposed; averaging more independent measurements of the same k-space location (or, more generally, acquiring more total data) improves the achievable SNR — the physical basis for the NEX/NSA (number of excitations/signal averages) parameter documented in the MRI Parameters cluster.
• Scan time scales with the number of phase-encode steps acquired. Because each phase-encode line requires a full repetition of excitation and readout (Section 3.4), and because finer resolution requires more phase-encode steps (first bullet above), a direct, physically-grounded three-way tension exists between spatial resolution, SNR, and scan time — every acceleration technique documented on MRIninja (parallel imaging, partial Fourier, non-Cartesian efficient trajectories) is, fundamentally, a strategy for partially escaping this tension by acquiring less raw k-space data than the naive Cartesian Nyquist requirement, at some identifiable cost (SNR penalty, artefact risk, or both) that must be weighed against the time saved.

8. MRI Technologist and Radiologist Pearls — Common Misconceptions

• “Each point in k-space corresponds to a point in the image.” The opposite is true (Section 5.1): each k-space point contributes to the entire image simultaneously, which is why k-space data corruption typically produces diffuse artefact (streaking, ghosting) rather than a spatially localised defect.
• “The centre and edges of k-space are equally important.” They serve distinctly different roles (Section 5.3): the centre predominantly governs contrast and SNR, while the periphery predominantly governs fine spatial detail — a distinction directly exploited by k-space-weighted and centric-ordered acquisition strategies.
• “Partial Fourier acquires exactly half of k-space.” In practice it acquires somewhat more than half (Section 6.3), specifically to allow correction of the real-world phase errors that make perfect conjugate symmetry an idealisation rather than an exact physical fact.
• “Non-Cartesian trajectories are just a different way of doing the same Cartesian scan faster.” Radial and spiral trajectories (Section 6.2) have genuinely different artefact behaviour, motion robustness, and reconstruction requirements from Cartesian sampling — they are not simply accelerated Cartesian equivalents.
• “Reducing scan time is ‘free’ as long as you keep the same resolution.” The resolution/SNR/scan-time relationship (Section 7.2) is a physical constraint, not a software limitation — any acceleration technique that reduces scan time while holding resolution constant necessarily does so at some identifiable cost elsewhere (typically SNR, sometimes artefact risk), even when that cost is well managed rather than eliminated.

9. Practical and Clinical Relevance

Every acquisition-time and image-quality trade-off documented across MRIninja’s anatomical protocol and parameter pages traces back to this page’s physics. The Parallel Imaging, NEX/NSA, and Acquisition Matrix parameter pages are all, at root, different strategies for manipulating how much of k-space is acquired and how (Sections 6–7). Motion artefact management strategies discussed in the anatomical protocol pages frequently rely, at a physical level, on exploiting the differential motion sensitivity of different k-space trajectories and sampling orders (Sections 6.1–6.2). Field-of-view and phase-oversampling decisions (covered practically in the MRI Parameters cluster) are direct applications of the frequency/phase-encode gradient physics described in Sections 3.3–3.4, specifically concerning what happens when anatomy extends beyond the encoded field of view along either axis.

11. Evidence Gaps and Ongoing Debate

• Optimal trajectory choice remains application- and hardware-dependent. No single k-space trajectory (Cartesian, radial, spiral, or echo-planar) is universally superior; the choice involves genuine, actively-researched trade-offs between motion robustness, off-resonance sensitivity, reconstruction complexity, and achievable acceleration, and different vendors and research groups continue to make different practical choices for similar clinical applications.
• Partial Fourier phase-correction methodology. As referenced in Section 6.3, multiple distinct algorithms exist for estimating and correcting the phase errors that limit theoretical partial Fourier time savings (homodyne reconstruction, POCS-based iterative methods, and others); no single approach has become universally standardised, and algorithm choice can materially affect image quality at aggressive partial Fourier factors.
• Non-Cartesian and highly undersampled reconstruction validity. As acceleration techniques increasingly combine non-Cartesian trajectories with partial Fourier and parallel imaging (and, increasingly, AI-based reconstruction), the general question of how to rigorously characterise image-quality and diagnostic-reliability trade-offs at aggressive combined acceleration factors remains an active area of technical research, more fully addressed in the planned Reconstruction and Post-Processing Physics child page group (master page Section 4.11) than here.

12. Evidence-Based References

A. Guidelines / Consensus / Society Recommendations

No dedicated society guideline exists for foundational k-space and image-formation physics as such — this is textbook/landmark-paper physics rather than a clinical practice area subject to society guidance. Category A is therefore not populated for this child page.

C. Important Prospective / Original Studies

D. Technical MRI Papers

End of document — Signal Localisation and Image Formation (k-Space) — Child Page under the MRIninja MRI Physics — Fundamentals and Principles master page — v1.0 — July 2026 Parent page: MRI Physics — Fundamentals and Principles