7.3 Reconstruction Algorithms — Filtered Backprojection vs. Iterative Reconstruction

Why Reconstruction Algorithms Are Tested

Image Reconstruction is ARRT's subcategory 1.D under Image Formation, listing eight named leaf items: filtered backprojection, iterative, prospective/retrospective, raw data versus image data, reconstruction algorithm, reconstruction slice thickness, reconstruction interval, and interpolation. This section covers the first five; Section 7.4 covers the remaining three parameter-focused items. Reconstruction is not confined to Image Production, either — ARRT's Safety specifications explicitly list "image reconstruction (e.g., iterative, retrospective, artifact suppression software)" as a named dose-minimization technique under Radiation Protection (1.B.1.g), so this content is tested twice: once as image quality, once as dose optimization.

Raw Data Versus Image Data

Raw data is the actual set of attenuation measurements ("projections") recorded by the detector array at every gantry angle during the scan — essentially unprocessed sinogram data. Image data is the final cross-sectional pixel/voxel image produced after a reconstruction algorithm has been applied to that raw data. The testable consequence: as long as the scanner retains the raw data, it can be reconstructed and re-reconstructed multiple times — different kernel, different slice thickness, even a different algorithm (FBP versus iterative) — without rescanning the patient and without any additional dose. Once image data is generated, it cannot be reversed back into raw projections; a technologist who needs a different kernel later still needs the raw data on file, not just the finished images.

Filtered Backprojection (FBP)

Filtered backprojection is CT's traditional analytic reconstruction algorithm. A mathematical filter — the convolution kernel — is applied to each projection to correct for blurring, and each filtered projection is then "backprojected" (smeared back across the image matrix along its original acquisition angle) and summed with every other projection to build the final image. FBP is fast and remains the quickest reconstruction option on most scanners, but its central limitation is that image noise rises predictably as dose falls — cutting dose in half roughly doubles apparent image noise with FBP alone — which caps how far dose can drop before an FBP image becomes clinically unusable.

Iterative Reconstruction (IR)

Iterative reconstruction describes a family of newer algorithms that build the image through repeated comparison-and-correction cycles: the algorithm compares a modeled/estimated image against the actual measured raw data (blending statistical noise modeling with the FBP framework, or, in more advanced "model-based" forms, modeling the entire scanner geometry and physics) and iteratively refines the image to reduce noise and artifact. The payoff: IR can produce a lower-noise image at the same dose as FBP, or let dose drop while keeping noise at an FBP-equivalent level. Published comparisons report roughly 30-50% dose reduction with statistical iterative techniques and up to ~60% with model-based iterative reconstruction (MBIR) at matched image quality — exact performance varies by vendor algorithm (e.g., ASIR/ASiR-V, SAFIRE/ADMIRE, iDose4/IMR, AIDR 3D) and the blend/strength level selected. Trade-off: pushing IR blending too high can produce an over-smoothed, "waxy" image texture some readers dislike, and IR takes meaningfully longer to compute than FBP.

Reconstruction Algorithm (Kernel Selection)

Independent of the FBP-versus-IR choice, the operator selects a convolution kernel that trades spatial resolution against noise. A smooth (soft-tissue/standard) kernel favors low noise and strong contrast resolution — appropriate for evaluating solid organs and soft-tissue detail. A sharp (bone or lung) kernel favors edge detail and spatial resolution at the cost of visibly higher image noise — appropriate for cortical bone detail or lung parenchyma. Because kernel selection is applied during reconstruction rather than acquisition, the same raw data set is routinely reconstructed twice from one acquisition: a chest CT, for example, is commonly reconstructed once with a soft-tissue kernel to evaluate the mediastinum and once with a lung kernel to evaluate the parenchyma — at zero additional patient dose.

Prospective Versus Retrospective Gating

Most directly relevant to cardiac CT (Chapter 12), gating strategy is tested here as a reconstruction-timing concept. Prospective gating ("step-and-shoot" cardiac acquisition) uses the patient's ECG signal to trigger x-ray acquisition only during a preselected, predictably quiet phase of the cardiac cycle — typically mid-to-late diastole — so data (and dose) are collected only for that one phase; it requires a slow, regular heart rate and yields just that single reconstructable phase. Retrospective gating instead acquires a continuous helical dataset across the entire cardiac cycle while recording the ECG trace alongside it, allowing the technologist or reader to reconstruct images at any cardiac phase after the scan is complete — valuable for arrhythmia or functional/ejection-fraction analysis — at the cost of substantially higher dose unless ECG-based tube current modulation reduces exposure during less useful phases.

Exam Scenario

A radiology department is transitioning its abdomen/pelvis protocol from FBP to a statistical iterative reconstruction algorithm to support a lower-dose CT screening initiative. The technologist expects roughly 30-50% less dose can be used while keeping image noise comparable to the prior FBP protocol, but knows to watch for an overly smooth, waxy image texture if the blend level is pushed too high. Separately, a cardiac CT patient with an irregular heart rate cannot be scanned with prospective gating (which needs a predictable diastolic window), so the technologist chooses retrospective gating instead, accepting a higher dose in exchange for the ability to reconstruct whichever cardiac phase turns out to have the least motion.

Key Takeaways

• Raw data (unprocessed projections) can be reconstructed repeatedly with different kernels, thicknesses, or algorithms at no added dose; image data cannot be reversed back into raw data.
• Filtered backprojection is fast but noise rises predictably as dose falls; iterative reconstruction reduces noise at a given dose (or allows dose reduction at matched noise), roughly 30-50% for statistical IR and up to ~60% for model-based IR, at the cost of processing time and possible over-smoothed texture.
• Reconstruction kernel (algorithm) selection — soft-tissue versus bone/lung — is independent of the FBP-vs-IR choice and can be applied twice to one raw dataset at no added dose.
• Prospective gating triggers acquisition only during one predictable cardiac phase (lower dose, needs a steady heart rate); retrospective gating acquires the whole cardiac cycle continuously (any-phase reconstruction, higher dose).