Artifacts: Recognition & Remediation

Artifacts are the physics speaking out loud

An artifact is any signal in the image that does not correspond to true spin density at the displayed location. Every artifact is a violation of one of the assumptions built into the imaging chain, the Larmor relation, uniform fields, a static object, adequate sampling, or a clean receive channel. Because each failure mode leaves a characteristic spatial signature, you can run the inference backward, reading the pattern to name the broken assumption, and then change the one acquisition parameter that restores it. This lesson treats artifact recognition as applied physics rather than a memorization list.

Two facts organize most of what follows. First, the phase-encode axis is acquired one line per repetition time (TR), separated by hundreds of milliseconds, while the frequency-encode (readout) axis is sampled in a few milliseconds within a single echo. Anything that changes between repetitions, motion, flow, pulsatility, instability, therefore corrupts the slow phase axis and propagates there. Second, both the frequency-encode position and the chemical-shift, susceptibility, and off-resonance displacements all live on the same readout axis, because that axis maps frequency to position. Keep those two anchors in mind and most artifacts become predictable.

Motion and flow: ghosting along phase encode

Consider an object that moves periodically with frequency $f_m$ (cardiac near 1 Hz, respiration near 0.2 to 0.3 Hz). Each phase-encode line is acquired at a slightly different motion phase, so the true k-space data are multiplied by a periodic modulation. Multiplication by a periodic function in k-space convolves the image with a comb of delta functions: the object is replicated into discrete ghosts. The ghost spacing in the phase direction depends on the ratio of the motion period to the TR-times-matrix acquisition time.

For a TR of 0.5 s and respiration at $f_m \approx 0.25$ Hz, ghosts recur every $0.25 \times 0.5 = 0.125$ of the FOV, giving roughly eight overlapping copies, a smeared blur rather than a single crisp ghost. Faster, lower-amplitude pulsatile flow tends to produce a few discrete, well-separated ghosts. Flow also carries spins that experienced an excitation pulse into the slice during readout, depositing bright or dark signal that maps to the wrong phase position.

Remedies attack the modulation directly. Cardiac or respiratory gating and breath-holding synchronize acquisition to a fixed motion phase, removing the periodic variation. Saturation bands placed over moving fat or inflowing vessels null the offending signal before it is encoded. Flow compensation (gradient moment nulling) rephases first-order motion so flowing spins return to baseline phase at the echo. And because ghosts always travel along phase encode, swapping the phase and frequency directions steers the ghost away from the anatomy of interest, for example moving CSF-pulsation ghosts off the spinal cord.

Aliasing: wrap-around when the FOV is too small

The Nyquist condition for the phase axis fixes the field of view through the phase-encode step size $\Delta k_y$. The reconstructed FOV is the inverse of that step, so any anatomy lying outside the FOV is not absent, it is undersampled and reappears folded onto the opposite side of the image.

Spins at position $y$ outside the FOV accumulate phase $\phi = \gamma G_{PE} y\,\tau$ that is indistinguishable from the phase of a spin at $y - \text{FOV}_y$ inside the FOV, because the two differ by an integer multiple of $2\pi$ at every sampled line. The discrete Fourier transform cannot tell them apart, so out-of-FOV anatomy aliases to the wrong side. Wrap is most common on the phase axis because that is where we shrink the FOV to save time, the frequency axis is usually oversampled for free by the receiver.

Remedies enlarge the effective sampled extent. Phase oversampling (no-phase-wrap) acquires extra phase lines spanning a larger FOV and then discards the wrapped border, preserving resolution at a scan-time cost. Frequency oversampling is performed automatically by digitizing faster than Nyquist requires. Surface-coil sensitivity that excludes distant anatomy, saturation bands over the wrapping tissue, and simply enlarging the FOV or reorienting the phase axis along the patient's narrow dimension all reduce wrap.

Chemical shift: two distinct artifacts on the frequency axis

Fat protons resonate about 3.5 parts per million below water because their electron shielding differs. In hertz this offset scales with field strength: $\Delta f = 3.5 \times 10^{-6} \times \gamma B_0 / 2\pi$, giving roughly 220 Hz at 1.5 T and 440 Hz at 3 T. Because the readout encodes position as frequency, this constant frequency offset is misread as a constant spatial displacement of fat relative to water.

Type 1 chemical shift is this spatial misregistration along the frequency axis. At 1.5 T with a per-pixel bandwidth of 110 Hz/pixel, fat shifts $220 / 110 = 2$ pixels; drop the bandwidth to 55 Hz/pixel and the shift doubles to 4 pixels, and at 3 T it doubles again. It manifests as a bright fat rim on one side of a structure and a dark gap on the other, classically at the renal poles, the orbital fat, and vertebral endplates. The corrective parameter is increased receiver bandwidth, which shrinks the shift in proportion, at the cost of signal-to-noise that falls as the square root of bandwidth. Lower field and fat suppression also abolish it.

Type 2 chemical shift is a completely different mechanism: signal cancellation, not displacement. In gradient echo the fat and water magnetization precess at different frequencies and drift in and out of phase. At the out-of-phase echo time the two vectors are antiparallel, so any voxel containing both fat and water loses signal, producing a black India-ink outline around every fat-water boundary, around organs nestled in fat and within fatty-marrow vertebrae.

At 1.5 T with $\Delta f \approx 220$ Hz, the first opposed-phase TE is $1 / (2 \times 220) \approx 2.3$ ms and the first in-phase TE is $1 / 220 \approx 4.6$ ms. At 3 T these halve to about 1.15 and 2.3 ms. Type 2 cancellation never occurs in a properly formed spin echo, because the 180-degree refocusing pulse rephases the chemical-shift phase difference at the echo. That is why opposed-phase imaging is a gradient-echo technique and why exploiting it, the loss of marrow or adrenal signal between in- and opposed-phase, is the basis of diagnosing microscopic fat in adrenal adenomas and hepatic steatosis.

Susceptibility: local field distortion, worst on gradient echo

Magnetic susceptibility $\chi$ describes how a tissue magnetizes in $B_0$. At interfaces between materials of differing $\chi$, air-tissue boundaries, hemorrhage, calcification, and especially metal, the local field becomes nonuniform. This off-resonance does two things: it dephases spins within a voxel, causing signal loss, and it adds a spatially varying frequency that is misencoded as geometric distortion along the readout, pile-up and stretching of anatomy. The induced field offset scales with $B_0$, so susceptibility artifact worsens at 3 T.

Gradient echo has no refocusing pulse, so static off-resonance dephasing accumulates over the full TE and signal decays at $T_2^{*}$, making GRE acutely sensitive, which is exactly why susceptibility-weighted imaging detects microbleeds. Spin echo applies a 180-degree pulse that reverses static dephasing, recovering signal and greatly reducing the artifact. The remedies follow: prefer spin echo or fast spin echo, use a short TE, increase receiver bandwidth to compress the per-pixel frequency-to-distance mapping, shrink voxels, and align the readout to spread distortion favorably. Near orthopedic hardware, dedicated metal-artifact-reduction sequences (view-angle tilting and slice-encoding for metal artifact correction, VAT and SEMAC) add extra encoding to recapture displaced signal.

| Artifact | Encode axis | Worse with | Primary fix | | --- | --- | --- | --- | | Motion / flow ghost | Phase | Long TR, pulsatility | Gating, sat bands, flow comp, swap PE/FE | | Aliasing (wrap) | Phase | FOV smaller than object | Phase oversampling, larger FOV, sat bands | | Chemical shift type 1 | Frequency | Low bandwidth, high field | Higher receiver bandwidth, fat sat | | Chemical shift type 2 | Voxel-wide | Opposed-phase TE on GRE | Use spin echo or in-phase TE | | Susceptibility | Frequency + signal loss | GRE, long TE, metal, 3 T | Spin echo, short TE, high BW, MARS | | Gibbs / truncation | Both | Low resolution, sharp edges | Larger matrix, apodization filter | | Zipper / spike | Across image | RF leak or bad k-space point | Shield RF, repair sample, reacquire |

Truncation (Gibbs) ringing and the finite k-space window

Real acquisitions sample k-space only out to a finite $k_{\max}$. Mathematically, the true spectrum is multiplied by a rectangular window, and the image is therefore convolved with the inverse transform of that rectangle, a sinc function. At any sharp edge, the sinc point-spread function produces oscillating overshoot and undershoot, Gibbs ringing, parallel lines decaying in amplitude away from the boundary.

The first overshoot reaches about 9 percent of the edge step regardless of matrix size, the Gibbs constant, but with a larger matrix the rings move closer together and become finer, so coarser matrices show more conspicuous ringing. The classic site is the spinal cord on sagittal images, where low-resolution phase encoding produces alternating bright and dark bands inside the cord that can mimic a syrinx. Remedies trade resolution for ringing: increase the matrix (raise $k_{\max}$) and apply apodization (k-space filtering) such as a Hamming window that tapers the edge of k-space, suppressing the sinc sidelobes at the cost of slight blurring.

Explore: how the k-space window itself creates artifacts

The two artifacts above, aliasing and Gibbs ringing, are not separate phenomena bolted onto MRI; they are direct consequences of how k-space is sampled. Use the explorer to make this concrete. Widen the sampling step $\Delta k$ (undersample) and watch the field of view collapse so anatomy folds onto itself, that is wrap. Then truncate $k_{\max}$ (keep only the central low-frequency lines) and watch sharp borders sprout parallel rings as the effective sinc point-spread function broadens, that is Gibbs.

Zipper and spike: corruption of the data stream

Some artifacts originate not in the spins but in the receive chain. A zipper is a band of alternating bright and dark dots crossing the image, usually from stray radiofrequency entering the bore, a leaking room door, an unshielded monitor or anesthesia cable, or RF feedthrough from the scanner electronics. A narrowband RF source is effectively a single frequency, which the readout maps to a single column or a line of points, hence the zipper stripe. The fix is environmental: ensure the RF shielded room (Faraday cage) is intact, the door is sealed, and no unfiltered cables penetrate the penetration panel.

A spike or herringbone artifact is the dual phenomenon in k-space. A single corrupt k-space sample, from a gradient arc, a static-discharge spark, or a loose connector, is a point in the frequency domain. The Fourier transform of a single off-center delta is a sinusoidal plane wave spanning the entire image, producing a regular crosshatch of light and dark stripes across every pixel. The remedy is to find and eliminate the source (replace the faulty component) or to detect and replace the outlier sample before reconstruction; simply reacquiring often clears a transient spark.

Putting it together on real volumes

Recognition is a clinical skill, so practice on actual data. Open the viewer and scroll the T1- and T2-weighted volumes. Inspect the orbital and renal fat margins for type 1 chemical-shift rims, watch the phase axis for any pulsation ghosts, look at air-tissue interfaces (paranasal sinuses, skull base) for susceptibility signal loss, and examine the cord and skin edges for fine Gibbs rings on the lower-resolution acquisition. For each candidate artifact, force yourself to name the encode axis and the single parameter you would change.