Radiologist

United States only, and the list counts marketing authorisations, not installations: a cleared device may be running in a thousand hospitals or in none. Most of these are 510(k) clearances, which turn on substantial equivalence to an existing device rather than a trial, and each authorisation covers one narrow intended use — a triage flag, a measurement, a detection aid — not reading a study. The FDA states the list is not comprehensive and was assembled primarily by matching AI-related terms in the authorisation summaries, so both the total and the radiology share are floors rather than counts. Panel is the lead review panel, not the specialty that uses the device.

A demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.

South Korea. The word that matters in Article 2(4)(ra) is use: the high-impact category covers the development and the use of medical devices and digital medical devices, so a hospital running a cleared AI reading tool is an AI-using operator under this Act and not merely a customer of one. Article 34(1) then attaches duties to that operator — a risk-management plan, an ability to explain the output and the main criteria behind it, documentation, and item 4, human management and supervision. Article 31(1) separately requires telling the user in advance that the service runs on AI. Two limits: the detail is left to a Presidential Decree not contained in this text, and the fundamental-rights impact assessment in Article 35 is only a best-efforts duty. This says nothing about how many radiologists a hospital employs, and nothing about accuracy — it fixes who has to be able to answer for the output.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

United States, and it measures entry to training, not employment: these are positions programmes chose to fund and offer, which is a bet on demand several years out rather than a count of work being done today. The figures are the PGY-2 line for Radiology-Diagnostic in Tables 3 and 8A — 997 (2022), 1,006, 1,017, 1,057, 1,083 (2026), filled by all applicants at 99.9%, 100.0%, 99.9%, 98.6% and 98.4% — the main entry point for the specialty; the smaller PGY-1 line rose 132 to 156 over the same period. A specialty can automate part of its work and still train more people, because volume of imaging and the supply of readers move independently; this record bounds the employment story, it does not settle what machines read.

Verifiable change in hiring, headcount, hours or job scope. Highest weight — but causal attribution still has to be argued, not assumed.

The United Kingdom only, and hospital radiology departments only. This is the College's 18th annual census of its own members, answered by clinical directors with a 100% completion rate, so its adoption figures are a count of a defined population rather than a sample. Two clauses carry the linked tasks. On answering for the machine, the report states that additional time would also be required by consultants to oversee and monitor the AI system itself, and prints a clinical radiologist's own words: it slows down my acute stroke reporting as I have to check the AI and I often conclude that it is incorrect, I then have to explain this in my report. On the report itself, AI tools to aid in drafting reports are in place in 11% of departments and, of those using them, 24% of clinical directors said the tools reduced their workload, which is the most positive figure for any use case in this census and still a minority. Overall adoption: 75% of departments use AI tools in some capacity and support for image interpretation is in place in 58%. Counter-signals from the same document, which are the reason it is worth recording: across every use case more than three-quarters of clinical directors report that AI tools are either having no impact or are increasing their workloads; AI implementation ranked last of seven productivity initiatives at 28% positive against 70% for more skill mix and administrative support; and the report concludes that at this stage it does not appear that AI tools in radiology are simultaneously enabling faster and more accurate reporting, detection of more disease, and a reduction in the backlog. What it does not establish: it counts departments that hold a tool, not studies a machine read, so it says nothing about what share of reading is machine-first; and the workload figures are clinical directors' judgements rather than measured time, while the adoption figures are counts. The College has a stake and it runs both ways, since it is campaigning for more training places and cites these findings in support, and it also recommends further government investment in AI tools. The date is derived from the census's own line that data collection began in September 2025; no collection end date is given and the report's copyright line reads June 2026.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

The federal classification regulation for radiological triage software, the category most AI tools that flag urgent findings are cleared under. It defines the device as software that notifies designated clinicians that time-sensitive images are available for review, and says it does not mark, highlight or direct users' attention to a location in the image, does not remove cases from a reading queue, and operates in parallel with the standard of care, which remains the default option for all cases; its special controls require makers to describe algorithm limitations and test triage performance. It governs one device category and does not cover detection or diagnostic software cleared under other rules.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

The Medicare payer's final rule. Under the 2014 statute, any clinician ordering advanced diagnostic imaging such as CT, MRI or PET must consult appropriate use criteria through qualified clinical decision support software. CMS says the statute's real-time, claims-based reporting requirement presents an insurmountable barrier to operationalising the program, finalises a pause of the program for reevaluation and rescinds its regulations, and says the expected savings are negligible because the program cannot be implemented as written; it adds that it expects clinical decision support to keep being used where it helps. The statute remains in force, the pause was for claims-processing rather than clinical reasons, and the program bound the ordering clinician, not radiologists.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

Radiologists at a cancer hospital evaluating commercial AI used to prioritise incidental pulmonary embolism on routine oncology CT scans, across three periods from April 2019 to September 2020: routine workflow, human triage, and AI-prioritised worklists. Of 11,736 scans in 6,447 patients, the software reached 91.6% sensitivity and 99.7% specificity; radiologists' missed rate of incidental embolism fell from 44.8% (47 of 105 scans) without AI to 2.6% (one of 38) with it, and time to notification fell sharply. The software looks for this one finding only; the vendor supplied it, and one author has ties to other AI vendors.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

The recording of the remark itself, published by the conference organiser, 84 seconds long. Read what he actually said rather than the version that circulates, because they differ in a way that matters: the widely repeated form is five years, and the recording continues with a concession the retellings drop — it might be ten years, but we got plenty of radiologists already. He also framed the claim as a prediction about capability, that deep learning would do better than radiologists because it could accumulate more experience, not as a claim about hospitals, employment or regulation. Two limits on this record itself. The words here were read from the video's auto-generated captions, which are a machine transcription, checked against the audio for the sentences quoted but not professionally transcribed. And a forecast is never evidence about work: this record moves nothing on this page, and sits next to the records of what happened — a regulator's authorisation count and a decade of training-post numbers — which is where a reader can do the arithmetic themselves.

A named person with standing publicly predicted something, on a date, in an attributable statement. It is recorded so that who said what, and when, stays checkable — and it never moves a task's assessment, because a prediction is not an observation. Its value arrives later: the record sits on the same page as the evidence about that occupation, so anyone reading the forecast reads the record of what happened next beside it. That is the reckoning; this site publishes no verdict on whether a forecast came true.

You are choosing a specialty that a very famous person publicly advised nobody to enter. Worth knowing why that advice did not hold: it was about one task, and the job turned out to be several — including signing for a miss, seeing what nobody asked about, and doing procedures. It is also worth knowing the honest risk: if cheap reading makes imaging cheaper, volume rises, and a job defended by volume is defended by something that can change.

The prediction being wrong is not the finding; the reason it was wrong is. It measured a task and concluded about a job. Apply the same test to your own week: the hours spent on studies where the question is narrow and the answer is usually normal are the exposed hours, and they are also the ones most easily counted by an administrator building a business case.

Stay and strengthen

Be the one who validates the tools

A model validated elsewhere is not validated here, and somebody local has to sign that it works on this population, on these scanners. Most departments have not named that person.

Reshape the role

Move toward the procedure

Interventional work is physical, irreversible and licensed to a person — this page judges procedures human-led, though that judgement is inferred rather than shown by a record.

Adjacent move

Go where the images are used, not read

Clinical informatics, imaging programme design and screening policy all need someone who understands what a reading actually establishes — and that is a scarce kind of person in rooms where those decisions get made.

This occupation is the reason to distrust that question. A decade ago a very well-known researcher publicly advised that training should stop, on the grounds that the machine would be better within five years — adding that it might be ten. On the narrow task he named, the machines did get very good. The job did not go away, because the job was several things and he had measured one. A signal you can check yourself, and a better use of the worry: of your last week, how many hours went to studies where the question was narrow and the answer was usually normal. That number is the exposed part, and it is also the number an administrator can count.

Wrong in its conclusion, and worth understanding rather than dismissing, because the mistake is the one everybody makes. It took a task — reading a high-volume, pattern-clear study — and concluded about a job. On that task it was directionally right. What it left out was who signs for a miss, who sees the finding nobody asked about, who declines the study that should not be done, and whether cheaper reading makes demand rise rather than fall. Dismissing it as hype teaches nothing; reading it as a category error tells you exactly what to check about your own work.

A clearance says a product may be sold; it does not say a hospital bought it, switched it on, or kept it on. This site keeps those three things separate on purpose, and the gap between them is where most predictions about this field have gone wrong. It is also worth knowing which direction the products point: most are narrow detectors and triage tools that change the order of a worklist, which changes when a person looks rather than whether one does.

This site does not tell anyone which specialty to choose, and a page that did would be worth less. What it can give you is the shape of the bet: the exposed half is the high-volume narrow read, the durable half is accountability, incidental findings and procedures, and the open question is volume — whether cheaper imaging means more of it. Ask that question of the people already doing it, and notice whether they answer with what their department actually measures or with what they hope.