7. CT Artifacts

The reconstruction mathematics of CT presumes a fiction: that the measured signal along every ray obeys a monoenergetic Beer–Lambert law, $I = I_0\,e^{-\int \mu(s)\,ds}$, so that the line integral $p = -\ln(I/I_0) = \int \mu(s)\,ds$ is a clean projection of a single attenuation field. Filtered back-projection (FBP) and, to a lesser degree, iterative reconstruction invert this Radon transform under the assumption of linearity. Physics-based artifacts are, almost without exception, the image-domain consequences of that assumption being false. The clinical x-ray beam is polyenergetic, the detector counts a finite number of quanta, and each voxel is a finite volume — three departures from the idealization that generate beam hardening, photon starvation, and partial-volume averaging respectively.

Beam hardening arises because the true measurement integrates over a spectrum: $I = \int I_0(E)\,e^{-\int \mu(s,E)\,ds}\,dE$. Because $\mu$ falls steeply with energy (the photoelectric term scales roughly as $Z^{3.x}/E^3$), low-energy photons are preferentially absorbed and the mean energy of the surviving beam rises — the spectrum hardens — as it traverses tissue. The effective attenuation per unit path therefore decreases with depth, so the log signal is a concave (sub-linear) function of true path length rather than the linear function FBP assumes. The scanner's water-based bowtie and reconstruction correction linearize this for a water-equivalent path, but residual error survives wherever the body departs from that calibration. Two signatures result. Cupping depresses central HU in a large uniform structure (the posterior fossa or a contrast-opacified bladder may read 10–20 HU below its periphery). Streaks and dark bands span the line connecting two high-attenuation structures — the classic interpetrous hypodensity (Hounsfield's dark band) between the dense temporal bones that can simulate a pontine infarct, or the dark bands flanking the contrast-filled subclavian vein. The expert reads these by their geometry: a hypodensity that is defined by the line between two dense objects, that respects no vascular territory, and that lacks mass effect is beam hardening, not infarction.

Photon starvation is a stochastic, not deterministic, failure: along the most attenuating rays (through the shoulders, the posterior fossa, a large pelvis, or metal) so few photons reach the detector that quantum noise dominates. Because the variance of the reconstructed value scales inversely with detected counts and the log transform amplifies noise at low signal ($\operatorname{Var}[p] \approx 1/N$ for $N$ counts), the high-attenuation azimuth contributes wildly inconsistent projections. FBP smears that variance into bright-and-dark streaks aligned with the direction of greatest attenuation. The distinction from beam hardening is mechanistic and actionable: starvation is a noise (variance) problem cured by more quanta — higher mAs, tube-current modulation, lower-noise kernels, or iterative reconstruction that down-weights low-count rays — whereas hardening is a bias problem cured by spectral correction.

Partial-volume averaging is the most pervasive and the most underappreciated. Each voxel reports a single number equal to the volume-weighted mean of the linear attenuation coefficients within it, $\bar{\mu} = \sum_i f_i \mu_i$ with $\sum_i f_i = 1$. A voxel straddling cortical bone ($\sim$1000 HU) and CSF ($\sim$0 HU) in equal fractions reports $\sim$500 HU, mimicking acute hemorrhage; a small lung nodule occupying half a thick voxel is diluted toward the surrounding $-800$ HU air and may vanish. The error is largest where attenuation gradients are steep and slices are thick, which is why thin-section ($\le$1 mm) acquisition with isotropic voxels is the single most powerful mitigation. The related partial-volume-induced pseudoenhancement of small renal cysts adjacent to enhancing parenchyma is a quantitative trap: a measured 25 HU rise need not mean a solid lesion. Expert practice is to confirm any density measurement on the thinnest available reconstruction, place the ROI on the most homogeneous portion of the structure, and treat HU values from voxels near a high-contrast interface as suspect.

Tomographic reconstruction assumes the object is stationary across the angular views that compose a slice. Each projection samples the same $\mu(x,y)$ from a different angle; the inverse Radon transform fuses them on the premise that they describe one consistent object. Patient motion violates this premise by making the views mutually inconsistent — the anatomy that generated the 0° view is not the anatomy that generated the 180° view — and the back-projector, unable to reconcile them, distributes the contradiction across the image as blurring, doubling (ghosting), and streak. Unlike physics artifacts, motion error is fundamentally a problem of temporal sampling relative to the velocity of the structure, and the governing quantity is the displacement of an edge during the temporal window contributing to a voxel: a structure that moves more than roughly one detector aperture during a half-rotation will register as misregistered.

Gross voluntary or involuntary motion (the agitated, encephalopathic, or pediatric patient; swallowing; ocular saccades; tremor) produces the most recognizable signatures. High-contrast edges — the skull's inner table, an air–soft-tissue boundary, a contrast-filled vessel — generate paired or curvilinear streaks tangent to the moving interface, and structures double or smear along the direction of displacement. The diagnostic tell is that the artifact emanates from a high-contrast moving edge and is unconstrained by anatomy. A subtler and more dangerous form in neuroimaging is when motion mimics pathology: streaks across the brainstem can simulate hemorrhage, and motion at the orbital roofs can fabricate or efface a subdural collection. Modern temporal resolution is bought by faster gantry rotation (sub-250 ms), and in the heart by half-scan reconstruction (using only $180° + \text{fan angle}$ of data so the effective temporal window is roughly half the rotation), multi-segment reconstruction that stitches the same cardiac phase from consecutive beats, and ECG gating that confines acquisition to diastasis. Cardiac motion that escapes gating produces the pathognomonic step/stair discontinuity at the interface between data segments and characteristic blurring of the right coronary and mid-left-circumflex territories, where a motion-blurred vessel can be over-called as low-attenuation (non-calcified) plaque or stenosis — a high-stakes false positive.

Respiratory motion is the dominant patient artifact in body CT and behaves differently because helical acquisition adds a second inconsistency: the table is translating while the chest wall and diaphragm move, so the z-interpolation that builds each reconstructed plane combines data acquired at different respiratory phases. The signatures are stair-step or banding artifact at the lung bases and diaphragmatic dome, blurring of basal vessels and nodules, and — most treacherously in CT pulmonary angiography — transient interruption of the contrast bolus by a deep inspiration of unopacified blood from the IVC, which can simulate filling defects in the lower-lobe pulmonary arteries. Respiration also degrades quantitative tasks: it corrupts coronary calcium scoring, low-dose lung-cancer-screening nodule volumetry (where a blurred margin inflates apparent volume and confounds growth assessment), and emphysema densitometry. Prevention is overwhelmingly the highest-yield lever — breath-hold coaching, the shortest feasible scan time (high pitch, wide detector, single-breath-hold or even single-rotation volumetric acquisition of the whole chest), and end-inspiratory or, for air-trapping protocols, paired end-expiratory holds. The Bayesian reading discipline is to demand corroboration: a basal "nodule" or a lower-lobe "embolus" that lies in the plane of maximal respiratory blurring, that lacks a convincing correlate on a second series, and that coexists with diaphragmatic banding should be attributed to motion and, when clinically material, re-imaged rather than acted upon.

Where physics artifacts arise from the beam and patient artifacts from the object, hardware artifacts arise from the instrument: high-density implanted material that overwhelms the detector and corrupts the projection data, and detector elements that drift out of calibration. Both are, at root, problems of inconsistent or missing data fed into a reconstruction that assumes the sinogram is complete and self-consistent.

Metal artifact is the superposition of every physics mechanism, driven to its extreme by an object whose attenuation lies orders of magnitude above tissue. A dental amalgam, hip prosthesis, spinal hardware, surgical clip, or pacing lead so attenuates the beam that, along rays traversing the metal, almost no photons reach the detector — severe photon starvation — while the surviving beam is profoundly hardened, and Compton scatter and detector saturation, undersampling, and edge effects compound the error. The sinogram entries behind the metal are therefore either missing (zero counts, dominated by noise) or grossly biased. FBP back-projects this corrupted data into the radiating bright-and-dark streaks and starbursts that obscure the peri-prosthetic bone, the spinal canal beside instrumentation, or the soft tissues around a dental restoration — precisely the regions of clinical interest (loosening, periprosthetic fracture, infection, recurrent tumor). The expert recognizes metal artifact by the dense object at the apex of the streaks and reasons about what the streaks hide: a dark band crossing the bowel adjacent to a hip prosthesis is not pneumatosis, and apparent lucency around a screw may be artifact rather than loosening. Severity scales with the metal's atomic number and density (steel and cobalt-chromium far exceed titanium), its bulk, and its orientation relative to the scan plane; a thin lead oriented along z may barely streak while a bulky transverse implant devastates the slice.

Ring artifact has an entirely different, almost diagnostic, generative geometry. In a third-generation rotate–rotate scanner each detector channel maps, over a full rotation, to a fixed radius from the isocenter; if a single channel is miscalibrated, defective, or drifting (gain error, bad reference, contaminated collimator), its consistent offset is back-projected into a perfect concentric ring (a full circle if the gantry rotates 360°, an arc for partial scans) centered on the rotation axis. The radius of the ring localizes the offending channel; the artifact's circular symmetry and its independence from anatomy distinguish it instantly from disease. Clinically, a faint ring can masquerade as a subtle pathology when it happens to overlie a structure — a ring through the brain can mimic a hypodense or hyperdense lesion, and rings in the abdomen can simulate a cystic wall — and a ring centered on the patient is the classic mimic of a true concentric structure. Because the cause is instrumental, the fix is instrumental: detector recalibration (air/water gain scans), flat-field correction, and on the modern scanner a software ring-correction filter applied in the sinogram (often a high-pass filter along the channel direction) or in polar image coordinates. Ring artifact is the canonical clue that a scanner needs service; persistent rings on multiple patients mandate physics/QA escalation rather than per-image post-processing.

Artifact management is a three-tier discipline — recognize, prevent, correct — and the expert deploys it in that order of priority, because a prevented artifact costs nothing and a recognized artifact at least costs no diagnostic error, whereas a corrected artifact always risks trading one error for another.

Recognition is fundamentally a Bayesian act of separating signal from instrument. Every candidate "finding" carries a prior probability of being disease versus artifact, and the radiologist updates that prior using features that disease cannot easily counterfeit: geometry (a hypodensity defined by the line between two dense structures is beam hardening; a perfect concentric circle is a detector ring; streaks radiating from a dense object are metal), respect for anatomy (artifact ignores vascular territories, organ boundaries, and fascial planes), reproducibility (true lesions persist across reconstructions, planes, and ideally phases, whereas motion and partial-volume effects shift or vanish on thin or reformatted data), and physical plausibility (a 500-HU "hemorrhage" at a bone–CSF interface, or a $-30$-HU "fat" focus straddling lung and soft tissue, is partial-volume averaging). The corresponding cognitive failure modes are as important as the physical ones. Satisfaction of search — stopping after the obvious finding and missing a second, real lesion lurking within an artifact-degraded region — is the dominant error around metal and at the motion-blurred lung bases. Anchoring and premature closure convert a plausible-but-artifactual appearance into a fixed (mis)diagnosis; framing by the requisition ("rule out stroke") biases a beam-hardening band toward being read as infarct. The discipline is to ask, before reporting any borderline finding, "what artifact could generate exactly this?" and to demand independent corroboration before letting it change management.

Prevention operates at acquisition and is the highest-yield tier. Patient cooperation and immobilization, breath-hold coaching, and the shortest feasible scan time (high pitch, wide-detector volumetric acquisition, fast rotation) defeat motion and respiratory artifact; ECG gating and half-scan/multisegment reconstruction defeat cardiac motion. Removing or repositioning external metal, angling the gantry to keep dense hardware out of the slice of interest, raising tube potential (higher kVp hardens the incident beam less relative to tissue and improves penetration through metal and large body habitus), and matching mAs and tube-current modulation to the attenuation profile prevent photon starvation. Thin collimation with isotropic voxels is the structural prevention of partial-volume averaging. Each preventive choice trades against dose, temporal coverage, or noise, and the expert balances these explicitly.

Correction is the post-acquisition tier, and modern CT offers a hierarchy matched to mechanism. Iterative reconstruction (statistical/model-based, IR/MBIR) down-weights low-count rays and incorporates a noise model, directly attacking photon-starvation streaks and reducing the dose needed to avoid them; deep-learning reconstruction (DLR) now delivers comparable noise suppression with more natural texture, though both can introduce a plastic, over-smoothed appearance and, in the case of DLR, rare hallucinated or erased structure that must be guarded against. Metal-artifact-reduction (MAR) algorithms treat the corrupted sinogram explicitly: they segment the metal trace, discard or in-paint the affected projections by interpolation from neighboring views or from a prior image, and re-reconstruct — effective for the dominant streaks but prone to new secondary artifacts (loss of fine peri-implant bone detail, spurious low-density bands, and fabricated edges) that can themselves mimic loosening or lucency, so MAR and non-MAR series are best read side by side. Spectral approaches are the most physically principled correction for beam hardening and metal: dual-energy CT synthesizes virtual monoenergetic images at high keV (typically 100–200 keV), where the photoelectric contribution and thus beam-hardening and metal streaks are markedly attenuated, and supports water/iodine and other material decompositions that also defeat pseudoenhancement. Photon-counting CT extends this further — its intrinsic energy resolution, near-elimination of electronic noise, and small native detector elements reduce beam hardening, suppress metal and (by calibration) ring artifacts, and improve the spatial/contrast resolution that mitigates partial-volume error — and is, as of 2026, the most comprehensive single-platform answer to the physics-based and hardware artifacts described in this chapter. The unifying principle is that no correction is free: every algorithm that fills missing data or suppresses noise is inventing information under a model, and the expert validates the corrected image against the uncorrected data and against clinical plausibility before trusting it.