The MRI Reinvention: Faster Scans, Better Images, and the Real Role of AI

Imagine trying to photograph a hummingbird with a camera that needs a very long exposure. You could ask the bird to hold still. You could build a more sensitive camera. Or you could take a shorter, noisier exposure and use clever software to work out what the full picture must have looked like.

That, in essence, is the story of magnetic resonance imaging (MRI) over the past decade. MRI produces some of medicine’s most detailed pictures of soft tissue — brain, joints, prostate, heart — but it is slow, loud and expensive. Artificial intelligence is now being used to make scans faster and images cleaner. It is a genuine advance. It also comes with a failure mode that deserves plain explanation.

Why MRI takes so long

As the U.S. National Institute of Biomedical Imaging and Bioengineering (NIBIB) explains, an MRI scanner uses a powerful magnet to align protons — mostly in the body’s water — with its magnetic field. Pulses of radio waves knock those protons out of alignment; as they relax back, they emit signals the scanner detects. Unlike CT scans, MRI does not use ionizing radiation (the kind of radiation in X-rays that can damage DNA).

The trade-off is time. The scanner does not take a snapshot. It gathers its measurements piece by piece in a mathematical space that engineers call k-space, then converts them into an image. More measurements mean a sharper picture but a longer scan — and patients must stay very still throughout, inside a machine that NIBIB notes can reach sound levels of up to 120 decibels. Long scans mean fewer patients per machine, more motion blur, and a hard time for children, people in pain and anyone who is claustrophobic.

Plain English: An MRI scanner doesn’t photograph you. It collects thousands of radio measurements and calculates a picture from them. Collect fewer measurements and the scan is faster — but the calculation gets harder.

Where AI actually enters the scanner

Engineers have long used tricks to skip measurements and fill the gaps mathematically, such as using multiple receiver coils at once. Deep learning — a type of AI that learns patterns from large sets of examples — adds a new approach: train a neural network on many fully sampled scans, and it can learn to reconstruct a good image from far less data.

Some of the clearest evidence comes from knee MRI, a common, well-standardized exam:

• In the fastMRI project, a collaboration between NYU Langone Health and Facebook AI Research, an AI model built images from four times less data. In a 2020 study in the American Journal of Roentgenology, six musculoskeletal radiologists reviewing scans from 108 patients found the AI-accelerated images interchangeable with standard ones for diagnosis.
• A later prospective study in Radiology — meaning patients were scanned with the new method in real time, rather than old data being reprocessed — enrolled 170 people referred for knee MRI. Deep-learning reconstruction cut mean scan time from about 9 minutes 56 seconds to 5 minutes 33 seconds, and the images were judged diagnostically equivalent.

Major scanner manufacturers now offer AI-based reconstruction on commercial machines. The FDA’s public list of AI-enabled medical devices shows radiology as the largest category, spanning reconstruction, image enhancement, triage and measurement tools.

The hallucination problem

There is a reason radiologists remain cautious. When a network fills in missing data, it is, in effect, making an educated guess about what the image should contain. Usually that guess is excellent. Sometimes it is not.

A widely cited 2020 study by Vegard Antun and colleagues, published in PNAS (preprint here), tested several deep-learning reconstruction methods and found that tiny, almost undetectable changes to the input could produce severe artefacts — and that small structural details, such as a small tumor, might fail to appear in the reconstructed image. Counterintuitively, adding training data could sometimes make performance worse.

An AI-reconstructed MRI can look perfect and still be wrong. The danger is not a blurry image — it is a convincing one.

This does not mean AI reconstruction is unsafe; the clinical studies above found diagnostic equivalence in their settings. It means validation has to be specific. A method proven for knees on one type of scanner is not automatically proven for small brain lesions, children, or unusual anatomy. The questions to ask about any AI-accelerated scan are practical ones:

1. Was it tested prospectively, on real patients, in the body region at hand?
2. Did radiologists compare diagnoses — not just image “quality”?
3. Were rare and subtle findings included, or only common ones?
4. Is there a way to check the original, unprocessed data when something looks odd?

The quieter reinvention: hardware, helium and contrast

Not all of MRI’s reinvention is software. Most clinical scanners use superconducting magnets that must be kept extremely cold with liquid helium. As the Radiological Society of North America has reported, helium is a finite resource with periodic supply disruptions, and manufacturers have been developing scanners that need much less helium, or none. Researchers and companies are also building lower-field and portable systems that trade some image quality for access — bringing scanning to places a conventional suite cannot reach.

Contrast agents are another frontier. Many MRI exams use gadolinium-based contrast agents, injected to make certain tissues stand out. In December 2017 the FDA required a new class warning after finding that gadolinium can remain in the body, including the brain, for months to years. The agency said retention had not been directly linked to adverse health effects in patients with normal kidney function. Reducing contrast doses — including with AI methods that enhance low-dose images — is an active research area, not yet a settled replacement.

The real role of AI: less magic, more plumbing

Strip away the headlines and AI’s contribution to MRI looks less like a robot radiologist and more like better plumbing:

• Reconstruction: faster scans from less data, the most mature use.
• Image clean-up: reducing noise and motion blur.
• Workflow: positioning patients, choosing scan settings, and flagging urgent findings for faster review.
• Measurement: automatically outlining organs or lesions so size changes can be tracked.

Each saves minutes; together they could let the same scanner serve more people. The open question is whether faster scans translate into better outcomes, or simply more scans. That depends on how health systems use the time saved.

What to watch

Look for prospective studies beyond joints — brain, prostate, heart and children’s imaging — and for evidence that accelerated scans catch subtle disease as reliably as standard ones. Watch too how regulators treat reconstruction software that keeps being updated after approval. For the parallel story in cancer screening, read Can AI Spot Cancer Earlier?, and for the wider hardware landscape, The Next Medical Device Wave. Want to see how MRI contrast and slices work? Try our MRI Explorer.