DoseSave
DoseSave · Topics · How Does CT Reconstruction Work? FBP, Iterative and Deep Learning — Animated
Reconstruction · Animated

How Does CT Reconstruction Work? FBP, Iterative and Deep Learning — Animated

A CT scanner does not produce a slice directly; it measures 'shadows' (projections) from hundreds of angles and computes the image back from them. We follow the three generations of this computation step by step, with animations: backprojection, filtered backprojection, and iterative/deep learning.

Intermediate· ~14 min read· Sourced

A CT scanner does not "take" a slice directly. All it does is measure how much the rays crossing the patient from hundreds of angles are attenuated. The image is computed back from those measurements — and that is reconstruction. Here we follow how that computation works, and its three generations, step by step.

Projection and sinogram

A measurement from a single angle is a projection: a profile giving the total attenuation of the rays in that direction — a kind of shadow.1 As the tube and detector rotate, a projection is collected from each angle; together they form the raw data called the sinogram.

one projection per anglesinogram (projections vs angle)Measurement → sinogram
The profile at each angle (a projection) is written as a row of the sinogram. A single point traces a sine curve there — hence the name.

Backprojection: why blurry?

The most intuitive idea is backprojection: each projection is "smeared" back across the image plane along the direction it was measured. But when these smears, summed over 360°, overlap, you get not a sharp point but a 1/r blur — so simple backprojection inevitably blurs the image.1

The smears overlapresult: blurry blob (1/r)true pointbackprojection →+ one angle
Every smear piles up at the center; overlapping smears turn a sharp point into a blob with a 1/r profile. That is why simple backprojection alone is not enough.

Filtered backprojection (FBP)

The fix is clever: pass each projection through a filter (convolution) before backprojecting it. The filter adds negative side-lobes to the profile that exactly cancel the 1/r blur, leaving a sharp image.1 This was the standard method for decades. The filter choice is a fundamental trade-off: a sharper filter raises spatial resolution but also raises noise.1

measured profilefilter ⊗filtered profile (negative side-lobes)unfiltered → blurryfiltered → sharpFBP: filter first, then backproject
The filter adds negative side-lobes; when backprojected, they cancel the neighboring blur and a sharp image remains. The filter choice sets the sharpness–noise balance.

Iterative reconstruction (IR)

Iterative reconstruction solves the problem in a loop. The start may be a guess (a constant image or an FBP image). Then: from the current image estimate, forward projections are generated and compared with the measured projections; the difference is the error matrix. The algorithm updates the image so as to reduce this error, and the loop repeats.1 It is computationally intensive; all major manufacturers have implemented it, and its main benefit is the potential for dose reduction.1

① Estimate(initial image)② Forwardprojection③ Compare withmeasured → error④ Update theimageas the error shrinks:error matrix → shrinksLoop: estimate → forward project → compare → update
Each pass forward-projects the estimate, compares it with the measured data, and updates the image to reduce the difference. After a few passes, noise can be lower and dose smaller.

Deep-learning reconstruction (DLR)

The newest generation is deep-learning networks that aim to suppress noise while preserving tissue texture, using patterns learned from high-quality reference images. Studies show DLR can improve lesion detection and low-contrast detectability even at low dose.23 Performance depends on the training data, architecture and clinical task, so local validation is essential.

low dose · noisyneural networknoise reducedDLR: learned denoising
The network cleans the noisy low-dose input using learned patterns. Its effect depends on protocol and task; in clinical use it must be validated with local quality control.
Related articles
For a table of each method's strengths and risks: FBP, Iterative and Deep-Learning Reconstruction. For CT dose metrics: CTDIvol, DLP & SSDE. For parameter effects: CT Parameters.

References

  1. Bushberg JT, Seibert JA, Leidholdt EM, Boone JM. The Essential Physics of Medical Imaging, 3rd ed. Lippincott Williams & Wilkins, 2011. Bölüm 10 (Computed Tomography) — projeksiyon (s.351), basit ve filtreli geri projeksiyon, 1/r bulanıklığı ve rekonstrüksiyon filtresi (s.353), iteratif rekonstrüksiyon (s.357). Atıflardaki sayfa numaraları bu baskıya aittir.
  2. Brady SL, et al. Improving Image Quality and Reducing Radiation Dose for Pediatric CT by Using Deep Learning Reconstruction. RadioGraphics, 2021. pubs.rsna.org
  3. Düşük kontrast ayırt edilebilirliği ve doz azaltımında derin öğrenme rekonstrüksiyonu — fantom çalışması. PMC8980706
Note: This content is for education; for clinical decisions or regulatory compliance, consult a qualified medical physicist and current regulations.

← All articles