BRENNEN

Industries / Health / Healthcare / Imaging centers

Reading images

Reading is the radiologist's work of looking through every image of a scan and writing the report that tells the referring doctor, and now often the patient, what it shows.

7

Milestones achieved, from 1949 to 2021.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1949

    Readers miss what is there

    The radiologist Leo Henry Garland first described in the journal Radiology what is now called perceptual error, when a reader fails to see an abnormality that is on the image, and proposed having two people read each image. A 2026 paper by Michael Bruno and colleagues at Penn State says the rate of such misses is essentially unchanged since then. Reported error rates run from 3 to 4% where most images are normal to about 30% in test sets with many abnormal cases. Many European breast screening programs came to use double reading, and in 2019 the European Commission's guideline group recommended two readers for each screening mammogram, a conditional recommendation on evidence it rated moderate in certainty.

    LessonBetter images did not lower the human miss rate, so the trade added checks such as a second reader.

    europepmc.orgeuropepmc.orgcancer-screening-and-care.jrc.ec.europa.eu

    Credit

    Leo Henry Garland

  2. 1993

    One vocabulary for mammograms

    The American College of Radiology (ACR) published the first edition of BI-RADS, the Breast Imaging Reporting and Data System. It gives a dictionary of terms for what a reader sees on a mammogram, a fixed order for the report, and final assessment categories, each tied to a recommended next step. The work began in the late 1980s because mammogram reports differed so widely. New editions followed in 1995, 1998, 2003 and 2013, the system has been translated into six languages, and it became the model for similar systems for the lung, liver, thyroid and other organs.

    europepmc.orgeuropepmc.orgacr.org

    Credit

    American College of Radiology

  3. 1993

    The film room closes

    The Baltimore VA Medical Center in Maryland moved its radiology department from film to screens, which its radiologists call the first filmless radiology department in the United States. Images were stored and viewed on a picture archiving and communication system (PACS), closely linked to the hospital's information and transcription systems. Eliot Siegel and Bruce Reiner reported clinical and economic gains and found that linking the PACS to the patient's electronic record was essential. As scanners produced more images per scan, the images to be read grew far faster than the scans: at Mayo Clinic in Rochester, Minnesota, CT and MRI scans doubled from 1999 to 2010, while the images per radiologist per minute of the working day rose from 2.9 to 16.1, one image every 3 to 4 seconds.

    europepmc.orgeuropepmc.orgeuropepmc.org

    Credit

    Eliot Siegel, Bruce Reiner and the Baltimore VA Medical Center; Robert McDonald and colleagues at Mayo Clinic for the workload count

  4. 2001

    Night reads from daylight

    NightHawk Radiology Services was founded in Coeur d'Alene, Idaho, to read US hospitals' night-time scans with American radiologists working in daylight in Sydney, Australia, and Zurich, Switzerland. By 30 June 2005 its 30 radiologists served 693 hospitals, about 12% of all US hospitals, mostly with preliminary reads of emergency scans, nearly nine in ten of them CT. The final read, the one insurers pay for, stayed with each hospital's own radiology group.

    sec.gov

    Credit

    NightHawk Radiology Services and its founders, among them Paul Berger

  5. 2002

    Scoring each other's reads

    The ACR piloted RADPEER at 14 facilities: when reading a new scan, a radiologist rates the earlier report of the same patient on a 4-point scale. Of 20,286 cases, 2.9% were disagreements in difficult cases and 0.8% were misreadings in cases that were not difficult, as confirmed by each practice's committee. In 2020 an ACR-sponsored summit at Harvard Medical School reported that score-based peer review had not been shown to improve radiologists' performance or to measure it validly, and called for peer learning instead: group review of cases for teaching, kept apart from performance evaluations. ACR accreditation now offers a peer learning pathway, and the ACR says a successor to RADPEER that combines both is coming.

    LessonA score kept on each radiologist was not shown to make anyone read better.

    europepmc.orgeuropepmc.orgacr.orgacr.org

    Credit

    J. P. Borgstede and colleagues, ACR; D. B. Larson and colleagues for the 2020 summit

  6. 2008

    Fewer typists, more errors

    Speech recognition, software that turns a radiologist's spoken report into text without a typist, has been used in radiology since 1981. Studies in three countries counted the errors left in final reports. At the University of Michigan, 59 of 265 final reports (22%) had significant errors, while most radiologists believed their own error rate was 10% or less. At hospitals in Toronto, 23% of 308 breast imaging reports made with speech recognition had at least one major error, against 4% of 307 typed by transcriptionists. At Tygerberg Hospital, a teaching hospital in South Africa, 9.6% of 300 such reports had clinically significant errors against 2.3% of typed ones, and four years later 24.3% of speech recognition reports still had errors of some kind.

    LessonRadiologists underestimated their own error rates, so the errors were seen only when someone counted them.

    europepmc.orgeuropepmc.orgeuropepmc.org

    Credit

    L. E. Quint and colleagues; S. Basma and colleagues; J. du Toit, R. Hattingh and R. Pitcher

  7. 2021

    Results before the doctor calls

    US information-blocking rules under the 21st Century Cures Act applied from 5 April 2021 and required test results, including scan reports, to be released to patients electronically without delay. A survey published in 2021, of 70 top US hospitals, found that 42% of those with patient portals held radiology reports back 4 days or more. A 2022 survey of 8,139 patients and care partners at four US academic medical centers found that 95.7% preferred immediate release, including 95.3% of those whose results were not normal; among those with abnormal imaging results, 17.5% reported feeling more or much more worried. England chose a delay: NHS England's March 2025 guidance, written with the Royal College of Radiologists (RCR) and the Society of Radiographers, tells hospitals to release imaging reports through the NHS App or other agreed channels four weeks after a report is formally verified.

    federalregister.goveuropepmc.orgeuropepmc.orgsor.org

    Credit

    J. L. Mezrich and colleagues (2021 survey); B. D. Steitz and colleagues (2022 survey); NHS England, RCR and Society of Radiographers

Watch for

  1. An open model for dictation

    Google published MedASR, a 105-million-parameter speech recognition model for medical dictation whose weights can be downloaded; its model card dates it 18 December 2025. Google reports a word error rate, the share of words transcribed wrongly, of 4.6% to 6.6% on its private test set of radiologists' dictation, against 25.3% for OpenAI's Whisper v3 Large, and a 58% relative cut in errors against Whisper on Eye Gaze, a public, credentialed-access set of chest X-ray dictations. Its model card says most training speakers were native English speakers raised in the United States and that it may do worse for others.

    Achieved whenan outside group measures MedASR on radiology dictation it was not trained on, including speakers with other accents, and publishes error rates close to Google's.

    developers.google.comarxiv.org

    Claimed by

    Google, December 2025 (paper May 2026)

Open

  1. 01

    A radiology dictation test

    A public set of recorded radiology dictations that anyone can download, with a scoring script that counts the errors that change a report's meaning: left and right swapped, a "no" dropped or added, a wrong number or unit, a wrong body part. At least 20 speakers, including speakers whose first language is not English, each read 50 de-identified or made-up reports, and at least four speech recognition engines are scored on the same audio.

    The brief
    Credit · unclaimed
  2. 02

    When patients see their reports

    An open, dated table of when patients in at least 40 countries or health systems can read their own imaging reports online: at once, after a set delay, or not at all, with the channel (a hospital portal or a national app), any exceptions such as cancer findings, and a link to the rule or policy behind each entry.

    The brief
    Credit · unclaimed

Open 01 · the brief

A radiology dictation test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of recorded radiology dictations that anyone can download, with a scoring script that counts the errors that change a report's meaning: left and right swapped, a "no" dropped or added, a wrong number or unit, a wrong body part. At least 20 speakers, including speakers whose first language is not English, each read 50 de-identified or made-up reports, and at least four speech recognition engines are scored on the same audio.

Why it is still open

The tests that exist are private or narrow. Google's radiology test set for MedASR is private. The one public set of radiology dictation, Eye Gaze, holds a single board-certified radiologist's reads of 1,083 chest X-rays and is open only to credentialed PhysioNet users. Appen's September 2026 MedTerm-90 benchmark scored seven engines by how many clinical terms survived in 90.5 hours of real dictation from 14 specialties, but its audio comes from Appen's licensed data library and is not public, and its summary page does not say which specialties are covered. The error studies of final reports measured reports, not engines, and found accents mattered in South Africa but not in Toronto or Michigan. GitHub holds small dictation apps built on MedASR or Gemini, but no shared test audio.

github.comdevelopers.google.comphysionet.orgappen.comeuropepmc.org

Built with what exists

report text, never images, from the Indiana University chest X-ray collection, whose 3,996 reports were de-identified and checked by hand before public release, or reports written for the test; volunteer readers recorded with consent; MedASR (downloadable from Hugging Face under Google's Health AI Developer Foundations terms), OpenAI's Whisper large-v3 (open weights) and ElevenLabs Scribe v3, whose "keyterms" let a developer list up to 1,000 words to listen for (the Note "Fewer word errors" covers Scribe v3); the jiwer library for word error rate; a short Python rule set for laterality, negation and numbers.

europepmc.org

  1. First step

    Write 50 short chest X-ray reports seeded with hard cases ("no evidence of", left and right, measurements in millimetres), record five volunteers reading them, and run Whisper and MedASR on the audio.

  2. Done looks like

    1,000 recordings (20 speakers, 50 reports each) from speakers of at least five first languages, released under an open license; for each engine, the word error rate and the number of meaning-changing errors per 100 reports, with 95% confidence intervals; two radiologists' agreement on what counts as meaning-changing, reported.

Open 02 · the brief

When patients see their reports

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open, dated table of when patients in at least 40 countries or health systems can read their own imaging reports online: at once, after a set delay, or not at all, with the channel (a hospital portal or a national app), any exceptions such as cancer findings, and a link to the rule or policy behind each entry.

Why it is still open

The published comparisons cover the United States only. A 2021 survey of 70 top US hospitals found practice varied widely, with 42% holding reports 4 days or more. Since then the US has required release without delay, NHS England's 2025 guidance set a four-week wait, and local services in England use 28 days (North East London) and 35 days (Essex). A search for an international comparison of release delays found none.

europepmc.orgsor.orgprimarycare.northeastlondon.icb.nhs.ukvirtualviews.essex.icb.nhs.uk

Built with what exists

health ministry, regulator and patient-portal pages; a public GitHub repository with one file per health system and a date checked; a browsing agent such as Browser Use (open source, MIT license; the Note "Deleting 80% to start over" covers it) to find candidate documents, with every entry confirmed by a person.

  1. First step

    Fill in the first 10 health systems, starting with the United States, England, Scotland, Wales and two Nordic countries, from official pages, and publish the template.

  2. Done looks like

    40 health systems on at least four continents, each with the delay in days, the channel, exceptions, a source link and the date checked; 10 entries rechecked by a second person, with agreement reported.