4. Image Reconstruction

Tomographic reconstruction is the inverse problem of recovering the spatial map of the linear attenuation coefficient $\mu(x,y)$ from a set of projections, each a line integral of attenuation acquired across many view angles. The forward model is the Radon transform, $p(s,\theta)=\int_{L(s,\theta)}\mu(x,y)\,d\ell$, and the foundational analytic inverse is filtered back projection (FBP). FBP follows directly from the Fourier slice (central section) theorem, which states that the one-dimensional Fourier transform of a parallel projection at angle $\theta$ equals a radial line through the two-dimensional Fourier transform of the object: $P(\omega,\theta)=\mathcal{F}_2\{\mu\}(\omega\cos\theta,\omega\sin\theta)$. Because back projection alone smears each projection uniformly along its ray and oversamples low frequencies (the $1/|\omega|$ density of radial sampling in Fourier space), reconstruction requires pre-weighting each projection by $|\omega|$ — the ramp filter — before back projection:

$\mu(x,y)=\int_0^{\pi}\!\!\Big[\int_{-\infty}^{\infty} P(\omega,\theta)\,|\omega|\,e^{\,i2\pi\omega s}\,d\omega\Big]_{s=x\cos\theta+y\sin\theta}\,d\theta.$

The ramp is mathematically obligatory, not a design choice, but it amplifies high-frequency noise without bound; every practical kernel therefore multiplies the ramp by a band-limiting apodization window (Shepp–Logan, Hann, or a vendor 'soft'/'sharp' family). This single product fixes the resolution–noise operating point and is the conceptual ancestor of every parameter discussed in the next section. FBP is linear, shift-invariant, exactly invertible for complete noiseless data, and computationally trivial — properties that make its noise and resolution analytically tractable and that, paradoxically, are exactly what nonlinear successors sacrifice.

FBP's weakness is that it treats every photon measurement as equally reliable and ignores the physics of detection. Iterative reconstruction (IR) reframes the task as optimization. Statistical iterative reconstruction minimizes a data-fidelity term in which each projection is weighted by its inverse variance — photon-starved rays through dense or large anatomy, which obey Poisson statistics with variance scaling as the inverse of detected counts, are down-weighted — typically with a regularizing prior $R$: $\hat{\mu}=\arg\min_{\mu}\;\tfrac12\lVert A\mu-p\rVert_W^2+\beta R(\mu)$, where $A$ is the system (forward-projection) matrix and $W$ the statistical weighting. Model-based iterative reconstruction (MBIR) additionally encodes the system optics into $A$ — finite focal-spot size, detector aperture, voxel geometry, and the polychromatic spectrum — and uses an edge-preserving prior (e.g., a Huber or $q$-generalized Gaussian potential) so that noise is suppressed in homogeneous regions while gradients are retained. The practical price is twofold: nonlinearity (resolution and noise become object- and dose-dependent, invalidating scalar metrics), the well-known 'plastic,' over-smoothed, or 'pixelated' low-contrast texture of strong MBIR, and long reconstruction times. Hybrid statistical algorithms (ASIR, iDose, SAFIRE, ADMIRE) were the clinical compromise that delivered 30–50% dose reduction with more familiar texture.

Deep-learning reconstruction (DLR) is the current frontier. Three architectural families dominate: image-domain convolutional denoisers trained to map a low-dose or FBP image to a high-dose target; dual-domain networks that operate jointly in the sinogram and image domains; and physics-informed unrolled networks that embed the forward model $A$ and a learned regularizer inside a finite number of unfolded gradient steps, marrying MBIR's data consistency to a data-driven prior. Trained on paired high- and low-dose exemplars, commercial systems (TrueFidelity/DLIR, AiCE, Precise Image, ClariCT.AI) achieve iterative-grade noise suppression at FBP-like speed and largely restore natural noise texture. Their distinctive and clinically consequential failure mode is that, being nonlinear and learned, they exhibit contrast- and dose-dependent spatial resolution — high-contrast edges are preserved while subtle low-contrast detail can be blurred or, in rare out-of-distribution cases, hallucinated or omitted entirely — which is precisely why image quality can no longer be certified by a single phantom number and must be assessed task-by-task.

The reconstruction kernel (filter) is the single most influential operator the interpreter controls after acquisition, because the same projection data can be reconstructed into a dozen images with radically different appearances. In analytic terms the kernel is the apodization applied to the obligatory ramp; in operational terms it sets the system's in-plane spatial resolution and, inseparably, its noise magnitude and texture. Resolution is quantified by the modulation transfer function $\mathrm{MTF}(f)$, the spatial-frequency response of the imaging chain, with the limiting resolution conventionally read at the 10% or 2% modulation points; sharp 'bone'/'lung' kernels push $\mathrm{MTF}(f)$ toward higher cutoff frequencies (finer detail), while smooth 'soft-tissue'/'standard' kernels roll the MTF off early. The catastrophe of the resolution–noise trade-off is that pixel noise scales with the area under the squared kernel transfer in frequency space, so sharpening a kernel raises high-frequency noise quadratically faster than it raises signal. A useful approximation for FBP is

$\sigma_{\text{pixel}}^2 \;\propto\; \frac{1}{N_0}\,\frac{1}{\Delta x^{3}\, h}\int |K(f)|^{2}\,df,$

where $N_0$ is incident fluence, $\Delta x$ the in-plane sampling, $h$ the slice thickness, and $K(f)$ the kernel — showing explicitly that resolution ($\propto$ kernel bandwidth), noise, slice thickness, and dose are coupled, never independent. Halving voxel dimension or slice thickness to gain resolution costs an eightfold or proportional increase in variance at fixed dose.

Because modern algorithms are nonlinear, a single MTF curve no longer characterizes them: in IR and DLR the effective resolution depends on local contrast and on dose. The field therefore measures the task transfer function (TTF) — an MTF analogue estimated from a finite-contrast edge (the circular-edge technique on a phantom rod) at a clinically relevant contrast and dose — together with the noise power spectrum (NPS), the Fourier decomposition of noise variance that captures not just magnitude but texture. A strongly regularized MBIR image and an FBP image can share an identical pixel standard deviation yet look completely different because their NPS peak frequencies differ; IR/DLR characteristically shift the NPS toward lower frequencies, producing the 'blotchy' or 'waxy' percept that can mask or mimic low-contrast lesions. AAPM Task Group 233 formalized exactly this TTF/NPS/object-dependent methodology as the standard for evaluating CT systems, precisely because scalar contrast-to-noise ratio became an inadequate predictor of clinical performance once reconstruction went nonlinear.

Noise optimization is consequently a multi-axis problem rather than a single dial. Vendors expose IR/DLR strength (e.g., DLIR-Low/Medium/High, ADMIRE 1–5), which trades noise suppression against potential low-contrast blur and texture distortion; the expert selects strength by task, favoring weaker settings where low-contrast detectability is paramount and texture fidelity matters (e.g., subtle hepatic lesions) and stronger settings where the task is high-contrast or where dose reduction is the priority. The table below anchors typical kernel families to their resolution-noise regime and intended task.

Kernel family	MTF 10% (lp/cm, approx.)	Relative pixel noise	Primary clinical task
Smooth / soft-tissue (B/Br low)	5–7	Low (baseline)	Low-contrast detection: liver, brain, mediastinum
Standard / body	7–9	Moderate (~1.3×)	General abdomen, routine survey
Sharp / bone (Br/Bl high)	12–15	High (~2–3×)	Cortical bone, fracture, temporal bone
Ultra-sharp / lung	13–18	Very high (~3–4×)	Bronchovascular detail, interstitial lung, HRCT

The interactive below lets you experience how a high-contrast, high-spatial-frequency target (metal electrodes and cortical bone) behaves under a sharp Bone window — the regime where kernel choice dominates and where reconstruction-domain artifact handling is most stressed.

Image quality in CT is not a scalar but a vector whose components — spatial resolution, contrast resolution, noise, and artifact burden — must be judged against a specified diagnostic task. Contrast resolution is the ability to distinguish structures of slightly different attenuation; it is governed by the intrinsic attenuation difference between tissues, by photon statistics (more photons sharpen the histogram and separate overlapping distributions), by the reconstruction algorithm's low-frequency noise behavior, and — at the point of perception — by windowing, since the displayed contrast is $\partial(\text{gray})/\partial(\text{HU})\approx 1/\mathrm{WW}$ over the window. Because CT digitizes attenuation onto the Hounsfield scale, $\mathrm{HU}=1000\,(\mu-\mu_{\text{water}})/\mu_{\text{water}}$, even a 10–20 HU difference between a hypodense metastasis and normal liver is recoverable only when noise is small relative to that difference and the window is narrow enough to map it across many display levels. The table situates the diagnostically critical low-contrast tasks on the HU scale.

Tissue / finding	Typical HU	Why it stresses contrast resolution
Acute infarct (gray–white loss)	20–30 vs 30–40	5–10 HU separation; defeated by noise or strong-IR blur
Hypovascular liver metastasis	20–40 vs 50–65 (liver)	~20 HU lesion-to-liver; window WW≈150 essential
Acute extra-axial hemorrhage	50–70 vs 30–40 (brain)	Modest contrast; subdural window widens conspicuity
Lipid-rich adrenal adenoma	≤10	Absolute HU is the diagnosis; demands accurate calibration
Cortical bone vs metal	~700 vs >3000	High contrast — limited by resolution/artifact, not noise

Signal-to-noise (SNR) and contrast-to-noise (CNR) are the classical figures, $\mathrm{CNR}=|\mathrm{HU}_A-\mathrm{HU}_B|/\sigma$, and they remain valid for linear FBP. Their fatal limitation under iterative and deep-learning reconstruction is that $\sigma$ measured in a uniform region does not predict the visibility of a structured lesion, because IR/DLR redistribute noise across spatial frequencies and bias resolution toward high-contrast edges. The modern, defensible figure of merit is the task-based detectability index $d'$, computed by a model observer that integrates the task function $W_{\text{task}}(f)$ (the frequency content of the lesion to be detected) against the system TTF and NPS:

$d'^{\,2}=\frac{\Big[\displaystyle\int |W_{\text{task}}(f)|^{2}\,\mathrm{TTF}^{2}(f)\,df\Big]^{2}}{\displaystyle\int |W_{\text{task}}(f)|^{2}\,\mathrm{TTF}^{2}(f)\,\mathrm{NPS}(f)\,df}.$

Because $d'$ rewards a reconstruction only for resolution and noise behavior at the spatial frequencies the lesion actually occupies, it correctly credits DLR's low-contrast detectability gains while exposing cases where headline noise reduction does not translate into better detection — and it is the metric on which dose-reduction claims should rest. A 40% noise reduction that leaves $d'$ unchanged for the relevant task offers no real-world benefit; conversely, matched-$d'$ comparisons are how 40–70% dose reductions are legitimately substantiated.

Artifact reduction is best reasoned about mechanistically, by acting on the generative physics in whichever domain it originates. Beam hardening, the preferential absorption of low-energy photons that makes the polychromatic line integral nonlinear in path length, produces cupping and dark inter-osseous bands; it is corrected by water/bone beam-hardening algorithms and, definitively, by spectral/monoenergetic reconstruction that synthesizes a single-energy dataset. Metal and photon starvation arise where counts fall to near zero and Poisson variance explodes, throwing bright-and-dark streaks; statistical reconstruction down-weights those rays and dedicated metal-artifact-reduction (MAR) methods inpaint or model the corrupted projections, though MAR can itself introduce new secondary artifacts — a classic failure mode the reader must anticipate. Ring artifacts trace to detector-element gain miscalibration and are addressed by gain normalization or sinogram filtering; cone-beam artifacts in wide-detector helical scanning are mitigated by exact/approximate cone-beam algorithms (e.g., Katsevich-type or advanced rebinning); and motion is suppressed by faster rotation, ECG gating, or motion-compensated reconstruction. The cardinal cognitive failure is satisfaction of search and the temptation to read through an artifact rather than recognize it: misreading a beam-hardening band as infarct, a MAR inpainting error as a real lucency, or DLR's smooth low-frequency noise as a genuine low-contrast lesion. Expert practice pairs each artifact percept with its physical cause and its domain-appropriate countermeasure, and treats any unfamiliar texture from a nonlinear algorithm as a hypothesis to be tested across windows and reconstructions before it is believed.