Industries / Health / Healthcare / Imaging centers
Machines
The scanners and supplies an imaging center runs on: X-ray rooms, CT and MRI scanners, mammography units, and the contrast agents injected so that blood vessels and organs show up.
7
Milestones achieved, from 1983 to 2022.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1983
X-ray pictures without film
Fuji Photo Film began selling computed radiography (CR) in 1983. A reusable plate takes the place of film, and a scanning laser reads the image stored in it out as numbers, so no darkroom is needed. The Osaka General Medical Center in Japan, rebuilding its radiology department at the time, used the system to make all of its general X-ray work digital.
Credit
Fuji Photo Film; M. Sonoda and colleagues described the system in Radiology in 1983
1990
A whole organ, one breath
Willi Kalender of Siemens Medical Systems in Erlangen, Germany, and colleagues moved the patient's table at a steady, controlled speed through a CT scanner that rotated continuously, so that the X-ray beam traced a spiral around the body. A computer then built images at any point along it, and a whole organ could be scanned in a single breath hold. A second team, led by Peter Vock, scanned the lungs of 24 patients this way and found that no single lung nodule was missed.
Credit
Willi A. Kalender (Siemens); Peter Vock and colleagues
2000
Mammograms go digital
On 28 January 2000 the US Food and Drug Administration (FDA) approved GE's Senographe, a full-field digital mammography system that records the breast X-ray on an electronic detector instead of film. In 2005 a trial that gave 49,528 women at 33 sites in the US and Canada both kinds of mammogram found the two equally accurate overall, but digital more accurate in women under 50 and in women with dense breasts. On 11 February 2011 the FDA approved Hologic's Selenia Dimensions system for tomosynthesis, a 3D form of mammography. By November 2026, 8,702 of the 9,140 US facilities certified to do mammograms had tomosynthesis units, by the FDA's count.
Credit
GE Healthcare; Etta Pisano and colleagues (the trial); Hologic
2019
AI cleans up low-dose CT
On 12 April 2019 the FDA cleared GE's Deep Learning Image Reconstruction, software that uses a trained neural network to turn a CT scanner's raw readings into a less noisy picture, so that a lower X-ray dose can still give a usable image. In a 2024 study at the First Affiliated Hospital of Zhengzhou University in China, 56 cancer patients had a standard scan and a scan at about half the dose (7.15 against 13.86 milligray, mGy) read this way. The half-dose scans showed all 88 lung nodules, but only 43 or 51 of 65 liver lesions 0.5 cm across or smaller, depending on the display setting, against 63 of 65 at the standard dose.
LessonA lower dose that works for one task can hide small findings in another, so each use needs its own test.
Credit
GE Healthcare; P. Hou and colleagues (the study)
2020
MRI at the bedside
On 6 February 2020 the FDA cleared Hyperfine's portable MRI scanner, which can be wheeled to a patient's bedside. Its magnet runs at 0.064 tesla, less than a twentieth of the 1.5 to 3 tesla of a standard hospital MRI. Yale New Haven Hospital has used it in routine care in its emergency department and intensive care unit since August 2020. When two neuroradiologists read 144 of its scans, they correctly found bleeding in the brain in 45 of the 56 patients who had it (80.4%).
Credit
Hyperfine Research; Kevin Sheth, Mercy Mazurek and colleagues at Yale
2021
CT that counts each X-ray
On 30 September 2021 the FDA cleared Siemens Healthineers' NAEOTOM Alpha, which the company calls the world's first photon-counting CT scanner. Its detector, a cadmium telluride crystal, measures each X-ray photon that reaches it, where earlier detectors first turn X-rays into light and measure the total energy of many at once. Laurel Burk of the FDA called it the first major new technology for CT imaging in nearly a decade.
accessdata.fda.govappliedradiology.comsiemens-healthineers.com
Credit
Siemens Healthineers
2022
One factory shut, scans cut
On 19 April 2022 GE Healthcare announced that its main factory for iodinated contrast, the iodine-based dye injected so that blood vessels and organs show on CT, had shut for a time during a COVID-19 outbreak in Shanghai, China. Its products Omnipaque and Visipaque ran short worldwide. The Radiological Society of North America (RSNA) reported that four companies supplied such contrast to the US and that the others could not quickly make up the gap. A review of 22 articles on the shortage found that hospitals in the US and Australia cut both the number of contrast scans and the amount of contrast given in each.
LessonA supply that rests mainly on one factory can slow scanning in many countries at once.
Credit
GE Healthcare (the supplier); D. R. Koeppel and I. B. Boehm (the review)
Watch for
Magnets that need little helium
A standard MRI magnet is kept superconducting by a bath of liquid helium, a gas in short supply. Siemens Healthineers says its Magnetom Flow.Ace 1.5 tesla scanner, cleared by the FDA in June 2025, needs 0.7 liters of liquid helium over its whole life in a sealed circuit with no quench pipe (the vent that lets the helium escape if the magnet warms suddenly), against more than 1,000 liters for conventional scanners. It also says the scanner uses over 30% less energy a year than its previous generation.
Achieved whenan outside survey of installed sealed magnets reports their helium top-ups, uptime and energy use over several years.
Claimed by
Siemens Healthineers, June 2025
Open
01
Map where contrast is made
A public table of every iodinated and gadolinium contrast product sold in the US, with each factory its label lists as making it, and a count of how many products, and which, depend on a single site. Extend it to other countries whose regulators publish manufacturing sites.
The briefCredit · unclaimed02
Hidden lesions, kept or lost
A test kit for AI software that rebuilds MRI images from fewer measurements, a way to shorten scans. The kit adds small synthetic lesions of known size, place and brightness to the raw scanner data of the public fastMRI collection, then records whether each open reconstruction model keeps, blurs or erases them, and whether it invents features in scans where none were added. Because fastMRI's terms forbid passing the data on, the kit is published as code and lesion settings that rebuild the same test from each user's own copy.
The briefCredit · unclaimed
Open 01 · the brief
Map where contrast is made
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every iodinated and gadolinium contrast product sold in the US, with each factory its label lists as making it, and a count of how many products, and which, depend on a single site. Extend it to other countries whose regulators publish manufacturing sites.
Why it is still open
In 2022 the shutdown of one factory in Shanghai caused a worldwide shortage. RSNA's report on it called for companies with large market shares to have "diverse production locations and capacity". Searches found commercial sourcing directories that list makers of the raw chemical, but no public map of which sites make which finished products.
Built with what exists
the DailyMed web service of the US National Library of Medicine, whose drug labels list each manufacturing establishment with its FDA establishment number and the product codes it makes; the label for Omnipaque, for example, lists GE Healthcare Shanghai, GE Healthcare AS and GE Healthcare Ireland; Python to parse the labels; a spreadsheet or small web page to publish the table.
First step
Pull the DailyMed labels for iohexol, iopamidol, iodixanol, iopromide and ioversol and extract their lists of establishments.
Done looks like
every iodinated and gadolinium contrast product on DailyMed listed with its manufacturing sites; the number of products, package sizes and companies that rely on one site; and a check against the 2022 shortage showing which listed products named the Shanghai site. Code and table public, rerun each quarter.
Open 02 · the brief
Hidden lesions, kept or lost
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test kit for AI software that rebuilds MRI images from fewer measurements, a way to shorten scans. The kit adds small synthetic lesions of known size, place and brightness to the raw scanner data of the public fastMRI collection, then records whether each open reconstruction model keeps, blurs or erases them, and whether it invents features in scans where none were added. Because fastMRI's terms forbid passing the data on, the kit is published as code and lesion settings that rebuild the same test from each user's own copy.
Why it is still open
The organizers of the 2020 fastMRI challenge, which drew 19 submissions on 7,299 brain scans, listed common failure modes across the entries. A 2026 study showed that small, noise-like changes to the input made two standard models, U-Net and E2E-VarNet, produce hallucinations, features that were not in the original, and that these "cannot be reliably detected using traditional image quality metrics". The nearest earlier work, by Francesco Calivá and colleagues in 2020, added small synthetic features to fastMRI knee data and found that a trained network left some of them out; they released code and a radiologist's marks of real knee lesions. Their features were made by attacking one network, so they are not a fixed set for scoring others, and a search found no such set.
Built with what exists
the fastMRI data from NYU Langone Health, raw de-identified k-space (the scanner's raw frequency data) from 6,970 fully sampled brain MRIs, available for research under a data use agreement; the fastMRI code repository, with PyTorch data loaders, undersampling functions and baseline models; Python with NumPy.
First step
Insert ten 5 mm lesions into one fastMRI brain case's raw data and confirm that a reconstruction from the full data shows all ten.
Done looks like
1,000 inserted lesions from 2 to 10 mm across 200 brain scans, plus 200 scans with none; for at least five open models, a published table of the share of lesions kept at each size and each speed-up factor, and the rate of invented features in the lesion-free scans.