CT Reconstruction (FBP vs Iterative)
- A CT scanner never "sees" a slice directly — it collects shadows (projections) from every angle, and reconstruction is the math that turns those shadows back into a picture.
- Filtered back projection (FBP) is the fast, classic recipe: smear each projection back across the image, with a sharpening filter so things don't blur into mush.
- Iterative reconstruction (IR) guesses an image, simulates what scanner it would have produced, compares it to the real data, and corrects — over and over.
- IR's big payoff is less noise at lower dose: you can dial down the radiation and let the algorithm clean up the grain.
- Push IR too hard and images get a waxy, plasticky, "oil-painting" look — different ugliness than FBP's streaks, but ugliness all the same.
Here's the uncomfortable truth about a CT scanner: it has never once actually looked at a cross-section of you. It can't. All it does is fire X-rays through your body from a thousand different angles and measure how much of each beam survived the trip — how much got eaten on the way through (that's attenuation). What comes back is a giant pile of shadows. Reconstruction is the magic trick that turns that pile of shadows back into the slice you stare at on the screen.
So the whole game is this: I have the shadows. How do I rebuild the thing that cast them?
The shadow problem
Imagine standing a friend in front of a wall and shining a flashlight at them. You get one shadow — a flat silhouette. Not enough to know their actual shape; a teapot and a doughnut can throw the same shadow if you pick the right angle. But now walk in a full circle, casting a shadow from every direction, and write down each one. With enough silhouettes from enough angles, you can mathematically reconstruct the 3D shape. That stack of angle-by-angle measurements is called the sinogram, and it's the raw food a reconstruction algorithm eats.
Filtered back projection: smear it back, carefully
The old reliable method is filtered back projection (FBP). The "back projection" part is intuitive: take each angle's shadow and smear it straight back across a blank image, in the direction the beam traveled. Do that for every angle and the smears pile up most densely where the real dense stuff was. A bright object starts to emerge from the overlapping streaks.
There's a catch. Plain back projection alone gives you a blurry, washed-out blob — like a photo with the contrast cranked all the way down. So before smearing, you run each projection through a sharpening filter (the "filtered" part) that boosts edges and cancels out the inevitable blur. Filter, then smear back, and a crisp image falls out.
FBP is fast, predictable, and has decades of trust behind it. For a long time it was simply how CT worked. Its weakness: it assumes the data is clean. When photons are scarce — low dose, or beams crawling through a big patient or a metal hip — FBP faithfully draws every bit of noise as ugly streaks radiating across the image.
Iterative reconstruction: guess, check, fix, repeat
Iterative reconstruction (IR) attacks the problem from the other end. Instead of one clever pass, it does detective work in a loop:
- Make a starting guess at the image.
- Mathematically simulate the projections that guess would have produced.
- Compare those fake projections to the real ones from the scanner.
- Nudge the image to shrink the difference, and go back to step 2.
It's like a sketch artist working a witness. Draw a face, ask "closer or further?", adjust, ask again — converging on the real culprit over many rounds. Because IR also carries a built-in model of how the scanner and the body actually behave (where noise comes from, the shape of the beam, sometimes the X-ray physics itself), it can tell signal from grain far better than a single sharpening filter can.
The headline benefit of IR is the noise-versus-dose trade. Because the algorithm suppresses noise so well, you can lower the radiation dose and still land at a diagnostic image — the reconstruction picks up the slack that fewer photons left behind.
Two flavors of ugly
Neither method is free of sin; they just fail differently. Knowing the failure signatures keeps you from misreading them.
| Filtered back projection | Iterative reconstruction | |
|---|---|---|
| Core idea | Filter each projection, smear it back | Guess an image, refine it against the data over and over |
| Speed | Very fast | Slower (heavier computation) |
| Noise at low dose | Streaky, grainy | Much cleaner |
| Characteristic artifact | Radiating streaks, prominent noise | Waxy, "plastic" or blotchy oil-painting texture if pushed hard |
Aggressive IR can make images look so smooth they feel fake — a waxy, plastic-skin or "oil-painting" texture where normal fine detail and even subtle lesions risk being smoothed away. Radiologists often run IR at a moderate strength rather than maxed out, precisely to keep the picture looking like a real human and not a wax figure. Reading at a setting you're not used to can fool your eye, so know which reconstruction you're looking at.
Why you should care at the console
This isn't trivia — it's dose and diagnosis. The move from FBP to iterative methods is a big reason CT doses have come down over the years: the algorithm absorbs the noise penalty of fewer photons. That same low-photon physics is what makes metal implants and beam hardening so nasty — more on that over in CT artifacts — and it's the same engine of measured attenuation that the Hounsfield scale is built on.
If you forget the rest, keep this: the scanner measures shadows, reconstruction rebuilds the slice, FBP does it in one fast filtered smear, and IR does it by guessing and correcting until the image stops arguing with the data — buying you cleaner pictures at lower dose, as long as you don't crank it until everyone looks like a candle.