Industries / Health / Healthcare
Imaging
Making pictures of the inside of the body with X-rays, sound, magnets and radio waves, and reading them.
7
Milestones achieved, from 1895 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1895
Seeing inside the body
Wilhelm Conrad Röntgen noticed crystals glowing behind a screen and traced the glow to a new kind of ray, the X-ray. One of his first pictures showed his wife Bertha's hand. He received a Nobel Prize in its first year, 1901. The date usually given for the discovery, 8 November 1895, is disputed.
Credit
Wilhelm Conrad Röntgen
1958
Echoes that show a mass
I. Donald, J. MacVicar and T. G. Brown reported in The Lancet that pulses of ultrasound, sound too high for people to hear, could be used to examine masses in the abdomen from the echoes they return.
Credit
I. Donald, J. MacVicar, T. G. Brown
1971
Slices through the body
In 1971 the radiologist James Ambrose and Godfrey Hounsfield, of the British company EMI, made the first CT examination of a patient. Computed tomography (CT) takes X-rays from many angles and has a computer build them into cross-sections. Scan time has fallen from about 5 minutes for a half turn to 0.24 seconds for a full turn. In 1973 Paul C. Lauterbur published a way to make images with nuclear magnetic resonance, the basis of MRI, which uses magnets and radio waves instead of X-rays. He and Peter Mansfield shared the 2003 Nobel Prize in Physiology or Medicine, and by then more than 60 million MRI examinations were done each year worldwide.
Credit
James Ambrose and Godfrey Hounsfield (CT); Paul C. Lauterbur and Peter Mansfield (MRI)
1993
One language for scans
Before a shared standard, it was very hard for anyone but a scanner's maker to decode or print its images. The American College of Radiology (ACR) and the National Electrical Manufacturers Association (NEMA) formed a committee in 1983 and published a first standard in 1985. Its third version, in 1993, was renamed DICOM (Digital Imaging and Communications in Medicine) and sent images over ordinary computer networks.
Credit
ACR and NEMA
2015
Software that did not help
The FDA approved computer-aided detection (CAD) for mammograms in 1998, and Medicare paid more for it from 2002. CAD then spread to most US screening mammograms, at a cost of over $400 million a year. A study of 625,625 mammograms read by 271 radiologists at 66 US facilities found sensitivity, the share of cancers caught, of 85.3% with CAD and 87.3% without. Specificity, the share of healthy women correctly cleared, was 91.6% and 91.4%.
LessonPaying for a tool before it is shown to help can spread it everywhere without benefit.
Credit
C. D. Lehman and colleagues, Breast Cancer Surveillance Consortium
2018
A diagnosis without a doctor
On 11 April 2018 the FDA authorized IDx-DR, software that examines photographs of the retina taken in a primary care office and decides on its own whether a person with diabetes has more than mild diabetic retinopathy, damage to the eye's blood vessels. In its trial of 900 people it caught 87.2% of cases and correctly cleared 90.7% of people without disease. The trial's authors call it the first FDA-authorized autonomous AI diagnostic system in any field of medicine.
Credit
IDx; M. D. Abràmoff and colleagues ran the trial
2026
AI as a second reader
In the MASAI trial in Sweden, 105,934 women were randomly assigned to standard reading of their mammograms by two radiologists, or to screening supported by an AI system. First results, for 80,033 women in 2023, showed 6.1 cancers found per 1,000 women against 5.1, the same false-positive rate of 1.5%, and 44.3% less reading work. Final results in 2026 counted interval cancers, those found between screens: 1.55 per 1,000 with AI against 1.76 without, meeting the trial's goal of being no worse. In Germany, the PRAIM study of 463,094 women at 12 sites found 6.7 cancers per 1,000 when radiologists chose to use AI, against 5.7 without.
Credit
Kristina Lång and colleagues, Lund University (MASAI); Eisemann and colleagues (PRAIM)
Watch for
Open models write X-ray reports
Google released MedGemma, AI models for medical images and text whose weights anyone can download. Google reports that in an unblinded study, a US board-certified cardiothoracic radiologist judged 81% of chest X-ray reports from MedGemma 4B accurate enough to lead to the same patient care as the original radiologist's report. A blinded study at Seoul National University Hospital, of five such models on 478 emergency patients, found four of them acceptable in 41.1% to 71.4% of cases, against 74.3% for radiologists' reports. It gave no separate figure for MedGemma.
Achieved whenan independent blinded study reports MedGemma's own acceptance and error rates at or near radiologists' level.
Claimed by
Google Research, July 2025
Drafting the whole report
Aidoc says its First Read software reads chest X-rays and drafts a report covering more than 100 set findings, which a radiologist must review and sign. The FDA gave it Breakthrough Device designation, which speeds review, in June 2026. It is not cleared for sale, Aidoc declined to list the findings, and no performance figures have been published.
Achieved whenthe FDA clears it and an outside study compares its drafts with radiologists' reports.
Claimed by
Aidoc, June 2026
Open
01
Blind-rate AI X-ray reports
A public, blinded rating of AI-written chest X-ray reports on images anyone can download. Run at least four report-writing models, including MedGemma, on the same 300 studies. Mix their reports with the original radiologist's report, hide which is which, and have three radiologists rate each report on the four measures used in the Seoul study: RADPEER (a standard scale for disagreement with a report), clinical acceptability, invented findings and clarity.
The briefCredit · unclaimed02
Find the follow-ups in reports
A labeled answer key and an open tool that reads radiology reports and lists every follow-up the radiologist recommended: the finding, the exam advised and the time frame. The tool is a paperwork aid. It reads report text only and makes no diagnosis.
The briefCredit · unclaimed
Open 01 · the brief
Blind-rate AI X-ray reports
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, blinded rating of AI-written chest X-ray reports on images anyone can download. Run at least four report-writing models, including MedGemma, on the same 300 studies. Mix their reports with the original radiologist's report, hide which is which, and have three radiologists rate each report on the four measures used in the Seoul study: RADPEER (a standard scale for disagreement with a report), clinical acceptability, invented findings and clarity.
Why it is still open
Google's 81% figure comes from one radiologist who knew which reports were AI-written. The one blinded comparison that included MedGemma used images from one Korean hospital that outsiders cannot get, and gave no figure for MedGemma alone. The public ReXrank leaderboard scores report writers with automatic text measures, not radiologists.
Built with what exists
the Indiana University chest X-ray collection, 3,996 de-identified reports with 8,121 images, free to download through the National Library of Medicine's Open-i service; MedGemma 4B and 27B, with downloadable weights; other models from the Seoul study whose weights are public; a simple web form for blinded rating. Because this collection is widely used, list which models say they trained on it and report those separately.
First step
Draw 300 studies at random from the Indiana collection, run MedGemma 4B on 20, and try the rating form with one radiologist.
Done looks like
300 studies, at least four models plus the original report, each report rated by three radiologists who do not know its source. Acceptance and invented-finding rates are published for each model, with confidence intervals and the agreement between raters, along with every report and rating so another team can rescore them.
Open 02 · the brief
Find the follow-ups in reports
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A labeled answer key and an open tool that reads radiology reports and lists every follow-up the radiologist recommended: the finding, the exam advised and the time frame. The tool is a paperwork aid. It reads report text only and makes no diagnosis.
Why it is still open
Recommended follow-ups are often vague or not done. At one Boston center, only 14% of follow-up recommendations held all the needed details before a software change raised it to 46%. In a sample of chest CT reports at another, referring clinicians acted on 13 of 33 recommendations for findings in the abdomen. Published extraction tools were built and scored on single hospitals' private reports, such as 1,009 CT reports from one hospital, and none of the studies found released its labels.
Built with what exists
MIMIC-IV-Note, which holds 2,321,355 de-identified radiology reports; access needs PhysioNet credentialing, CITI research-ethics training and a signed data use agreement. PhysioNet forbids sending this data to online AI services that keep it, are trained on it or have it reviewed by people, and recommends models run on your own machine, such as MedGemma 27B. Label Studio, an open-source labeling tool, for the answer key. The Note "Seven more flagged" covers a Claude-based reviewer that found issues in drug-making records; a hosted model like that may be used on this data only if it meets PhysioNet's conditions.
First step
Complete the CITI training and apply for PhysioNet credentialing, then write the labeling rules on 100 reports.
Done looks like
2,000 reports labeled by two people, with their agreement published. On 500 reports held back from development, the tool finds at least 95% of recommended follow-ups and gets the exam and time frame right for at least 90%. The labels are released on PhysioNet under the same credentialed terms, and the code is public.