DoseSave
DoseSave · Topics · FBP, Iterative & Deep-Learning Reconstruction
Reconstruction

FBP, Iterative & Deep-Learning Reconstruction

The path from raw projection data to a diagnostic image has passed through three major generations in 50 years. We examine how each works and its potential for dose reduction.

Intermediate· ~14 min read· Sourced

A CT scanner does not produce an image directly. As the gantry rotates, the tube and detectors record attenuation profiles from thousands of angles; turning this raw data into a slice image is called reconstruction. The core tension: when you lower dose by reducing photons, projections become noisy, and the algorithm's job is to cope with noise while preserving true anatomy.

Filtered Back Projection (FBP)

FBP has been dominant since the early 1970s. It is analytical: each projection is filtered and back-projected onto the image plane. It is fast, predictable and produces a natural noise texture. But at low dose, photon scarcity causes pronounced noise and streak artifacts.1

Iterative Reconstruction (IR)

Iterative reconstruction solves the problem in cycles: an image estimate is made, virtual projections are generated, compared with the real measurements, and updated until the difference shrinks. This yields less-noisy images than FBP from the same data, opening the door to the same quality at lower dose.2

The hidden cost of IR
Aggressive IR — especially below a certain dose threshold — can shift noise texture, producing a "plastic" look and reduced low-contrast detectability.

Deep-Learning Reconstruction (DLR)

The newest generation, DLR, aims to suppress noise while largely preserving tissue texture, using patterns learned from high-quality reference images. The training data, architecture and clinical task can vary by vendor and algorithm. Studies show DLR can improve lesion detection versus IR and FBP; in a phantom study where dose was reduced to a very low CTDIvol level (e.g. ~0.5 mGy), FBP detectability was compromised while DLR maintained it.3

Method Strength Risk Clinical note
FBP Fast, predictable, natural noise texture Noise and streak artifacts at low dose Still an important reference method
IR Reduces noise, can enable dose reduction Plastic look, reduced low-contrast detectability Aggressive use requires validation
DLR High noise-reduction + texture-preservation potential Depends on training data, algorithm and task Must be validated with local protocol and QC
Important caveat
DLR performance depends on its training data and the clinical task. Task-based validation and local quality control before clinical use are essential.

References

  1. Zeng D, et al. Fast filtered backprojection algorithm for low-dose CT. PMC8294203
  2. Brady SL, et al. Deep Learning–based Reconstruction for Lower-Dose Pediatric CT. RadioGraphics, 2021. pubs.rsna.org
  3. Low-contrast detectability and dose reduction using deep learning image reconstruction — 20-reader liver phantom study. PMC8980706
Note: This content is for education; for clinical decisions or regulatory compliance, consult a qualified medical physicist and current regulations.

← All articles