What happens to your image after you press scan
· 14 min read
Ask a radiographer what they would change about their scanner and the answer comes back in the same shape every time. More time. Less noise. Thinner slices. A patient who holds still. You can have some of that. You cannot have all of it at once, and no amount of clever software has changed that.
What has changed is where the argument gets settled. For most of the history of CT and MRI, the fix for a noisy image was to go back and collect more of it. More mAs. More averages. More phase-encoding steps. Reconstruction sat at the end of the chain doing the same job every time, and nobody spent much thought on it.
It does not work like that now. Iterative reconstruction, parallel imaging, compressed sensing and deep-learning recon have turned the back half of the chain into something that actively shapes what you see. Siemens sells Deep Resolve for MRI. GE, Philips and Canon all have deep-learning reconstruction for CT. They are built differently and marketed differently, and they are all chasing the same thing.
The question stopped being how do we collect more data. It became how much can we get out of the data we can realistically collect.
If you work a list, that cuts both ways. Recon can save a scan that would have been a repeat five years ago. It can also make a thin, rushed acquisition look perfectly respectable on the monitor when it is nothing of the sort. Telling those two apart is now part of the job.
Why you cannot just turn everything up
CT is easy enough to state. Image quality is bought with photons, and if you spend fewer you get more noise. It scales with the inverse square root of the dose, which is worse than most people's intuition: halve the dose and noise goes up by about forty per cent, not twenty. Nothing on the console repeals that.
MRI charges you in time instead. Smaller voxels mean less signal out of each one. More slices, more contrasts and fuller k-space coverage all lengthen the exam. And a long exam costs you things that never appear in a physics equation. The patient shifts. The breath-hold fails. The eleven o'clock slot eats the eleven-thirty. Somebody starts wondering whether sedation is easier than a third attempt.
Those are the walls. Everything below is an attempt to shift them a little.
From filtered back projection to something that learned
Filtered back projection is where everyone starts. Fast, direct, and it produces the grain radiologists have been reading for forty years. Its weakness is that it holds no opinion about noise. Hand it a low-dose scan and it reconstructs the noise every bit as faithfully as the anatomy.
Iterative reconstruction goes at it in a loop instead. It makes a guess at the image, works out what that guess would have measured, compares it against what the detector actually recorded, corrects, and goes round again. Model-based versions bring knowledge of noise behaviour and system geometry into the loop. You end up with less noise at the same dose, or the same noise at a lower one.
You pay in texture and in time. Push IR hard and images take on that smoothed, slightly plastic look plenty of readers dislike, and the heavier model-based versions can leave you waiting on the recon.
Deep-learning reconstruction is the industry's answer to that complaint. Each vendor describes its own work: GE says TrueFidelity uses a deep neural network trained on high-quality, low-noise FBP data. Canon says AiCE uses a deep convolutional network to tell signal from noise. Philips describes Precise Image as supervised learning built on a convolutional network. Different rooms, same target. Get the noise down without losing the texture and the low-contrast detail readers depend on.
At the console
"AI reconstruction" is not one thing. What the network trained on, what it trained to produce, where it sits in the pipeline, how it was validated: all of it turns up in the image on your monitor. Two systems can use the same phrase and give you visibly different grain.
So treat the recon setting the way you treat kVp and pitch. It is a protocol decision. It belongs in the protocol conversation, not in whichever preference happens to be saved on the console from the last shift.
Deep Resolve, and the problem it is pointed at
MRI got to the same place from the other side. Siemens describes Deep Resolve as AI-powered reconstruction meant to improve the balance between scan speed, resolution and image quality, and is careful to put it alongside parallel imaging and simultaneous multislice rather than in place of them. The reported wins are things like much shorter brain protocols, and finer in-plane resolution inside a scan time you can actually book.
The thing to hold on to is that this is not a filter running over a finished image. It sits inside the reconstruction itself, which is a much bigger question:
Can you take fewer measurements, or smarter ones, and still end up with an image somebody can report?
Philips makes a similar case in MRI with SmartSpeed Precise, which it describes as AI-based reconstruction for faster scans and sharper images. Different architecture, different training, same bargain underneath.
Parallel imaging came first
MRI has been accelerating scans since long before anyone put a neural network near a scanner. SENSE and GRAPPA use several receiver coils, each seeing the patient a little differently, so you can skip phase-encoding steps and work out what is missing from the coil geometry. That buys time. It costs SNR, which falls with the square root of the acceleration factor and then falls again through the g-factor. The g-factor is the part that catches people out. It gets worse the harder you push R, worse still where coil sensitivities overlap, and it tends to show up as noise right in the middle of the image where you least want it.
Learned reconstruction does not cancel that bill. It pays some of it back, which lets you sit at acceleration factors that used to be unusable.
Simultaneous multislice
Instead of exciting one slice at a time, SMS excites several together and unpicks them during reconstruction. In diffusion and functional work the time saving is substantial. It is the same trade as everything else here: a harder acquisition handed to a cleverer reconstruction.
Compressed sensing
Compressed sensing goes further and gives up on regular sampling altogether. It collects a deliberately incomplete, incoherent set of measurements, then reconstructs by leaning on two things it knows about medical images: they are not random, and one frame of a dynamic series looks a lot like the next. It has earned its place in cardiac, dynamic, 3D, angio and free-breathing work.
Push it too far and it leaves fingerprints. Over-smoothed images, a blocky look, fast motion smeared across frames. Which is why almost nobody runs it on its own any more.
Denoising is a decision, not a blur
The easiest way to make an image look better is to blur it. A conventional smoothing filter does exactly that, and it cannot tell a noise speck from a small low-contrast lesion, because across a handful of pixels there often is no difference. It takes both.
A trained network can do better, because it has seen a great deal of anatomy and can use context a local filter never gets to see. That is the real advance. It is also the real risk. A model that knows what anatomy usually looks like can hand you something convincingly anatomical in a spot where there was not much signal to work from.
The wrong objective
"Make it look cleaner." Anyone can produce a clean image. It tells you nothing about whether what you needed is still in there.
The right objective
Take out the noise, keep the findings, and be able to show that low-contrast detectability held up rather than just that a noise number went down.
Which is why judging a recon on noise alone is a trap. Standard deviation in a uniform patch of phantom will drop obediently for almost any smoothing you apply to it. Low-contrast detectability and task-based assessment are what actually track whether the study still answers the question.
Resolution and noise are the same dial
Resolution never comes from reconstruction alone. Detector element size, gradient performance, trajectory, slice thickness, interpolation, super-resolution methods and recon all have a hand in it. GE has pushed its deep-learning CT work toward high-resolution imaging. Canon positions AiCE for detail preservation, including ultra-high-resolution CT.
The part that matters for protocol work is the coupling. Every step toward finer detail costs signal: thinner slices, sharper kernels, smaller voxels, all of it. Decide one without the other and you get a protocol that is beautifully sharp and unreadable.
Artefacts follow the same rule. Some are best avoided at acquisition. Some want a physics-based correction. Some respond to a learned method. And some are solved by going back into the room and moving the patient. Not everything needs a neural network.
CT artefacts
MRI artefacts
At the console
The quickest way to learn where that coupling bites is to break it on purpose. Halve the mAs and find out which part of the image goes first. Sharpen the kernel and see how much of that sharpness you keep. Step the acceleration factor up one at a time until the g-factor noise turns up in the middle of the brain.
On a department scanner that experiment costs table time, and in CT it costs dose. In a simulator it costs nothing, which is why it is worth doing fifty times instead of reading about it once.
Image quality is a chain, not a setting
Here is the shift worth taking back to the department. Image quality is not decided at any one point any more. It builds up, and it leaks away, across the whole chain.
Fix one stage and ignore the rest and you leave a lot on the table. The best recon on the market will not rescue a coil that was not seated properly, a slice group planned off-axis, or a breath-hold instruction the patient did not understand.
Different names, same problem
The branding manages to make this field look more fragmented than it is and more interchangeable than it is, both at once.
| Manufacturer | Technology | Modality | Stated objective |
|---|---|---|---|
| Siemens Healthineers | Deep Resolve | MRI | Faster acquisition, higher resolution, image-quality optimisation |
| GE HealthCare | TrueFidelity DL | CT | Deep-learning reconstruction with FBP-like texture and lower noise |
| Philips | Precise Image | CT | AI reconstruction for noise reduction and low-contrast performance |
| Canon Medical | AiCE | CT | Noise reduction with spatial-detail preservation |
| Philips | SmartSpeed Precise | MRI | AI-assisted acceleration and image sharpness |
Each of these lives inside a different acquisition architecture, trained on different data, aimed at a different target image, validated in a different way. So the useful question is not whose AI is best. It is this:
How well does the whole system, acquisition and reconstruction and workflow together, turn what you captured into something a radiologist can act on?
The job got bigger, not smaller
There is a comfortable assumption that the cleverer reconstruction gets, the less the person at the console matters. Watch it play out in a department and you see the opposite. Recon raised what a good acquisition can turn into, which makes a good acquisition worth more than it used to be. It also brought in new ways for things to go wrong that only someone who understands the chain will spot.
Three things sit in a radiographer's working knowledge now that were optional ten years ago.
- What each recon setting does to the picture. Not the architecture. Just what it does to texture, to low-contrast detail, and to how hard you can push a protocol before it stops being worth running.
- What acceleration looks like when it fails. Residual aliasing, g-factor noise through the centre, slice leakage from SMS, the waxy look of compressed sensing pushed too far. They all have a signature, and spotting one is quicker than repeating the sequence and hoping.
- When a good-looking image should not be trusted, because the acquisition under it was thin and the recon did more of the work than the measurement did.
None of that arrives off a lecture slide. It is the sort of knowledge you build by changing something, seeing what happens, and doing it again. On a department scanner those repetitions compete with patients, and patients win. That is the gap ScanFlo was built for: the console workflow, the parameter physics, the artefacts and the dose readout, in a browser, where a wrong answer costs nothing.
Where this is heading
The direction of travel is toward systems where acquisition and reconstruction stop being separate decisions. Protocol chosen around the clinical question. Sampling adapted to the patient in front of you. Recon tuned to the task. Artefact correction handled quietly. Post-processing done before anyone asks for it.
Whether that turns up as one intelligent system or as a hundred small features over a decade, the effect on the floor is the same. Fewer protocols you can run on autopilot. More decisions that rest on knowing what the study is for. Protocols are easy to teach. Judgement is not. You build it one repetition at a time.
The measure has not moved
Image optimisation has grown from one step at the end of the chain into a stack of acquisition physics, mathematics, acceleration, machine learning and post-processing. Deep Resolve is what that looks like in MRI, where the scarce thing is time. The CT products from GE, Philips and Canon are what it looks like where the scarce thing is dose.
None of it is really a story about one vendor. It is about giving up on collecting more and getting better at using what you can collect. And the ceiling on that is still set at the console, by whoever planned the slices.
What counts as success has not changed: better clinical information, delivered consistently, at a quality clinicians can trust.
A note on sources. Vendor technologies here are described using the manufacturers' own published wording, and this article does not compare their clinical performance. The charts are schematic: they show the shape of a relationship, not measured values. Recon settings, dose levels and acceleration factors should always follow your department's protocols and local regulations.
Frequently asked questions
Does AI reconstruction let me simply lower the dose?
It widens the range of dose settings that still give you a diagnostic study, which is not the same thing. Noise still rises as dose falls; better reconstruction makes the rise gentler. Where the acceptable floor sits is a protocol decision, made against the clinical question and your local regulations, not something to infer from how clean the image looks.
What is the g-factor, and why does noise appear in the middle of the image?
The g-factor is the extra noise penalty parallel imaging pays on top of the square-root-of-R loss, and it depends on how well the coil sensitivities can be separated. Where those sensitivities overlap the separation is poorly conditioned, so the amplified noise concentrates centrally — which is exactly where you are usually looking.
Is deep-learning reconstruction the same as a denoising filter?
No. A smoothing filter works on a finished image and cannot tell a noise speck from a small low-contrast lesion, because over a few pixels there often is no difference. Learned reconstruction sits inside the reconstruction itself and brings context a local filter never sees. That is the advance and also the risk: a model that knows what anatomy usually looks like can produce something convincingly anatomical where there was little signal.
Why is noise standard deviation a poor way to compare reconstructions?
Standard deviation in a uniform phantom patch drops for almost any smoothing you apply, so it rewards blurring. Low-contrast detectability and task-based assessment measure whether the study still answers the question, which is the thing you actually care about.
Do the different vendor technologies do the same thing?
They are aimed at the same problem and built differently. Each sits inside a different acquisition architecture, is trained on different data, targets a different reference image and is validated its own way. Two systems can use the same phrase and give you visibly different grain, which is why the recon setting belongs in the protocol conversation.