Radiology 2066: a probabilistic forecast of the radiologist workforce

The short answer

Reading tips: "§" links go to the method behind a claim, superscript numbers show the source when you hover or tap them, and a "Back to where you were" button appears after any jump.

Probably not. Today's shortage most likely eases into rough balance, and the chance of at least 10% more radiologists than the work needs stays modest: about in 2045 and in 2055, mostly in futures with transformative AI. This relies on one assumption with thin evidence: that in a surplus, work flows back to radiologists from other doctors, as it seems to have after 2013. If it doesn't, the 2045 figure is (why).

The next decade. The market will most likely be close to balance: slightly short until about 2030, slightly over after that. In 2035, meaningful oversupply () is more likely than a meaningful shortage (), but both are unlikely.§

Why. Imaging keeps growing, but demand for radiologists grows more slowly, because AI absorbs much of the growth. In the median future in 2045:

With no further radiology AI, radiologist work would be higher than today
An older population, more imaging per person and bigger studies.§
AI saves of radiologist time
Drafting and measurement first, then AI-first reading of simpler exams as regulation and payment allow (details).
New work AI creates gives back about of that saving
So demand for radiologists ends up above today, including the market adjustment in step 4. AI is unlikely to create more radiologist work than it saves (why).
Supply grows faster, from today to , catching up with demand, so the median future is near balance ()
When supply and demand drift apart, the market closes part of the gap each year: work shifts between radiologists and other doctors, and labor-saving tools spread faster or slower, as in 1995–2025 (how we know). The market can absorb a gap of up to about of demand, so a 10%+ surplus needs supply to outrun demand by roughly 30% first. Ordinary AI progress, partly offset by growing imaging and new work, rarely gets there; transformative AI can.§

The risk rises until about 2055 and then levels off ( in 2055, in 2066). Nearly all of it comes from AI, most of it from transformative AI (how much). A renewed meaningful shortage is also possible ( in 2045, in 2066), mostly where imaging per person grows fast.

What it means. People entering practice in the next decade will most likely start in a market close to balance. If a surplus comes, it would most likely be felt as lower pay relative to other physicians and fewer openings for new graduates, not layoffs (pay and hiring). Even in balanced futures, radiologists' pay premium over other physicians likely shrinks, by about by 2045, as today's shortage premium fades. By career stage →

How sure. Trust the direction more than the digits. The percentages show what follows from our assumptions; they are not odds with a track record. Under other defensible assumptions the 2045 figure ranges from to (details).

Key terms

Medians, ranges and P(X)

Every number comes from the of simulated futures not already contradicted by events. The median is the middle value; the 80% range holds the middle 80% (10th to 90th percentile). P(X), or "chance of X", is the share of futures in which X happens in that year.

Demand and supply (FTEs)

Demand: full-time-equivalent radiologists (FTEs) needed to read the imaging patients and doctors want from radiologists, given the AI in use and after the market adjustment (during a shortage demand drifts down toward supply; during a surplus, up). Supply: practicing radiologists × the hours each works. Both are indexed so 2026 demand = 1.00, the work of about full-time radiologists at today's hours; about practice today, so supply ≈ .

Supply ÷ demand

The single best summary of the job market: below 1 is a shortage, above 1 a surplus. We estimate about today (80% range –), combining pay, vacancies and workload with a reconstruction of the job market since 1995; no measured national figure exists.

Meaningful oversupply and shortage

Supply ÷ demand above 1.10 (at least 10% more radiologists than the work needs) or below 0.90, a convention for a clearly noticeable imbalance. For scale, our reconstruction puts the 2012–13 market, when new graduates struggled to find jobs, at about , so meaningful oversupply is about twice that surplus. It is a pressure indicator, not an unemployment rate.

Market adjustment

In effect it pulls supply ÷ demand back toward 1, but only so far. When radiologists are scarce, other doctors take on more of the reading and labor-saving tools spread faster; the model assumes the reverse when radiologists are plentiful. Each year about of a gap closes (80% range –), a speed fitted to radiology's job market since 1995. History shows the market can absorb a gap of at least about 10% of demand; our central limit of about (range –) is partly judgment, and the evidence that surpluses are absorbed rests mostly on one recovery, 2013–2018 (details). Demand figures include it.

AI-assisted vs AI-first reading

AI-assisted: the radiologist reads; AI drafts, measures and flags. AI-first: AI reads and reports first; a radiologist audits, signs or handles escalations, which in the main model still takes about a fifth of the usual time on average (spot checks, escalations and sign-off, not a full review of every case). AI time saved combines both (saving a third of the time = 1.5× as much work per hour).

AI-progress regimes

How fast general AI improves is the largest unknown, so each future falls into one of four regimes: Stall ( of futures: radiology AI takes 1.6–3 times as long as under Trend), Trend (: today's pace), Fast (: about twice as fast), and Transformative AI (: AI eventually does nearly all of a radiologist's cognitive work). The weights are judgment calls, shown after removing futures already contradicted by events; change them in the explorer.

Optional: put yourself on the timeline

I am currently a…

Year by year, 2026–2066

Each chapter follows the median future; the panel beside it shows the full range for that year and updates as you scroll.

2026 · Today

A specialty that was supposed to be automated is short of people

In 2016 Geoffrey Hinton, a pioneer of deep learning, said we should stop training radiologists because AI would soon outperform them. Ten years later, diagnostic-radiology programs offered a record 1,241 positions and filled 97.6% of them. Average pay rose 6.6% in a year, and workforce projections still show a shortage a decade out.

AI has not gone away. More than 1,600 AI devices are FDA-listed, about three-quarters in radiology, but few are used at scale. Measured time savings range from none to 42% depending on the task and the study.§

Starting point: supply ÷ demand ≈ (we estimate a shortage of about ), AI saving almost no radiologist time yet, and about practicing radiologists.

2030

AI writes the first draft

In the median future, AI-drafted reports, automatic measurements and protocoling tools (which choose how each scan is done) save about of radiologist time. That is well below the one-third that a leading task-based analysis calls the five-year potential, because adopting tools takes years even after they work.§

Measured against 2026 demand, demand is and supply , and the shortage has mostly eased: the median supply ÷ demand is (medians don't divide exactly). Demand dips slightly because AI saves about of time and, while radiologists are scarce, some work shifts to other doctors. Unlike workforce projections, which leave out AI and assume work cannot shift, the model includes both. Meaningful oversupply in 2030 is likely.§

AI that reports normal chest X-rays on its own (tier 1) is already approved in Europe. In the median future the FDA authorizes such AI-first reading around 2029, but payment and liability acceptance take until the early 2030s.§

2035

Roughly balanced

AI now saves a median of radiologist time and reads of interpretive work AI-first, mostly tier-1 normal radiographs and negative screens. Demand is and supply ; the median supply ÷ demand is .

Meaningful oversupply is likely. Most of that risk is in the transformative-AI regime, where AI saves about half of radiologist time by 2035; outside it, oversupply is likely.§

2045

Near balance in most futures; a clear surplus mainly with transformative AI

In most futures, AI-first reading of all radiographs and screening exams (tiers 1–2) has cleared validation, the FDA, liability and payment.§ AI saves a median of radiologist time. The job has less dictation and more consultation, procedures and checking AI output.

The median supply ÷ demand, , is near balance, because the market adjustment absorbs moderate gaps. Meaningful oversupply () is more likely than a meaningful shortage () and comes mostly from transformative AI. How a surplus would be felt: pay and hiring.

2055

Complex reads begin to be automated

In the median future, AI-first reading of most CT and MRI (tier 3) becomes payable in the early 2050s. The AI-first share is , with a very wide range (–).

Median supply ÷ demand is . Meaningful oversupply is likely and a meaningful shortage .

2066

A different job

The median AI-first share is , yet AI saves only about of radiologist time: interpretation and report drafting are about 60% of the job, and AI-first reads still need audit or sign-off.

The market absorbs moderate gaps. Where AI moved very fast, the gap is too large to absorb, and residency programs cut back only after several years, so supply overshoots. Meaningful oversupply is likely. The forecast cannot say which future will happen; it says how likely each is and what to watch.

What it means at each career stage

When you enter practice determines which part of the forecast matters to you. Pick your stage above to highlight your row.

*Assumes a 1-year fellowship (most U.S. radiologists do one); subtract a year without it, and add one for each gap or research year. Each column is the chance in that one year. Oversupply is near zero for anyone starting before about 2030, because the model starts from a shortage. Today's shortage is smaller than the 10% cutoff for "meaningful", so meaningful shortage at the start is rare. Each figure can move by Later figures can be several times higher or lower under other defensible assumptions (How sure?).

Already practicing? Your horizon is the years to retirement. About 2.5% of radiologists retire or leave each year, so a slow fall in demand is absorbed by hiring fewer new graduates. Demand falls faster than that 2.5% a year at some point between 2035 and 2066 in of futures, nearly all with transformative AI. But new graduates keep arriving, so oversupply can grow even when demand is flat: severe oversupply (25% or more too many radiologists) lasting five or more years at some point before 2066 occurs in of futures (this counts any stretch over 40 years; in a single year such as 2045, severe oversupply is ).

How the job changes

What the job involvesfollows most directly from our task assumptions
Drafting, measurement and protocoling shrink; consultation, procedures and AI oversight grow (see How a radiologist's day changes).
Workload intensitylikely while short
In shortage futures, time saved becomes more studies per hour rather than shorter days.

Skills that are harder to automate, by assumption. The model assumes procedures (median AI time-saving ceiling ≈7%) and consultation (≈24%) are the least automatable tasks; this is an input, not a finding.

How a surplus or shortage would be felt: pay and hiring

This section translates the supply ÷ demand forecast into pay and hiring, using how radiologists' pay moved in past shortages and surpluses. It is not a separate forecast of wages. Radiologists' pay rose faster than other physicians' while they were scarce (2001–2006, 2022–2025) and fell behind while they were plentiful (2009–2014). Fitted to 2001–2014, that pattern is consistent with the 2022–2025 pay surge, but only after we recoded the earlier pay data, and it rests on a few mild imbalances, so treat these figures as rough.§

Radiologist pay relative to other physicians

Ratio of radiologists' pay to other physicians' pay, 2026 = 1. Median with 50% and 80% ranges. Shows the effect of shortage or surplus, including today's premium fading; not fee changes.

By 2045, in the median future, radiologists' pay relative to other physicians is about lower than today, as the premium built up during today's shortage erodes. It is at least 10% lower in of futures, and in the worst tenth about times today's relative level; it is at least 10% higher in . Pay falls in most futures because today's shortage premium fades, and a lasting surplus of any size pushes it down further.

A job market for new graduates at least as weak as 2012–13, when hiring was flat and new graduates struggled to find jobs, has a chance in 2045. In other AI-exposed jobs, early-career employment fell first. For scale, absorbing a 15% surplus by hiring alone would mean roughly halving new-graduate entry for about years; by hours alone, about fewer hours each.

Who keeps AI's time savings is a fee question the model does not settle, and the chart leaves it out. If fees per exam stay fixed, saving of the time per exam lets a radiologist earn about more per hour; if Medicare and insurers cut fees in proportion, payers keep the gain.

Why: AI time savings and the regulatory pipeline

Tiers of exam difficulty. Interpretive work is split into four tiers that AI could take over one after another: 1 normal radiographs and negative screens (≈7% of interpretive work), 2 other radiographs and screening (≈17%), 3 most CT, MRI, ultrasound and nuclear medicine (≈47%), 4 the hardest remaining work (≈29%). The grouping is our own, not a standard classification (why), so we also test automation in any order.

Being able to read a study is not the same as being allowed and paid to read it. For AI-first reading of each tier, the model chains: technical capability → clinical validation → FDA authorization → liability and payment → hospital adoption → actual time saved.§

What can today's AI do compared with radiologists, and what can't it? Two good overviews: a review of AI image interpretation in the New England Journal of Medicine and an accessible essay on why AI has not replaced radiologists. The full report summarizes the 2024–2026 studies.

AI time saved

Share of radiologist time that AI removes

AI-first share

Share of interpretive work where AI reads first

From capability to time saved, by tier of exam difficulty

Dot = median year each stage is reached; line = 80% range.

Anchors: autonomous normal-chest-X-ray AI approved in the EU in 2022; AI could safely self-report ~8–18% of chest X-rays; MASAI, a Swedish randomized trial of AI-supported mammography screening, took ~5 years from enrollment to its main result; no autonomous radiology read is FDA-authorized yet; jurors penalize radiologists who overrule AI; mammography computer-aided detection spread within years of Medicare payment.

How a radiologist's day changes

Share of radiologist working time by task, mean across futures. This follows most directly from our assumptions about which tasks AI can do.

Why: the Jevons question

Could cheaper, faster AI reads create so much new imaging that radiologists end up busier? Economists call that a Jevons paradox: making something more efficient raises total use of it, as with coal in 19th-century Britain. In the median future, new work offsets about of the time AI saves, and exceeds the saving in only of futures by 2045.§ That depends on new uses nobody can yet measure, such as image-derived biomarkers: with three times as many new imaging uses and twice the new radiologist tasks, the chance that new work exceeds the time saved rises from to (report §8.4).

How much of the saving is offset?

New work ÷ time saved: median and 80% range. Above 1 = a true Jevons paradox.

Why cheaper reads don't create much demand

The radiologist's fee is only ~20–25% of an exam's price, and imaging demand barely responds to price (a 10% cut in the patient's out-of-pocket price raises use by about 2%, and most of a fee cut would fall on payers, not patients). So halving interpretation cost adds only about 1–2% more exams.

The bigger channels are new uses (opportunistic screening on CTs done for other reasons, raising lung-screening uptake from 18%), new radiologist tasks, faster scanners and faster turnaround. AI also adds some radiologist work, such as oversight and follow-up of incidental findings, which the model includes. Technologist shortages cap the extra exams.

Why: the market adjusts

Training a radiologist takes years, but radiology's job market has repeatedly closed shortages and surpluses within a few years: other doctors read more while radiologists were scarce, and labor-saving tools spread faster. Version 1.6 adds this market adjustment and tests it against radiology's own history.

Testing against radiology's own history

We reconstructed radiology's job market from 1995 to 2026 from published growth in imaging, productivity and the workforce, and compared it with eight documented episodes, such as the mid-1990s glut, the 2000 shortage, the 2012–13 surplus for new graduates and today's shortage, each recorded at the time in job ads and workforce surveys. To make the test fair, we tuned the model only on 1995–2013, then checked whether it predicted 2015–2025 without having seen those years (we did set the input ranges in 2026, knowing what happened, so the test is not perfectly blind).§

Supply ÷ demand, reconstructed, against the documented job market

Each box is the range of supply ÷ demand consistent with what was documented at the time. Grey: the earlier model, with no market adjustment. Blue: with the market adjustment. Solid boxes were used for tuning; dashed boxes were kept aside as the test. Hover over a box for the probability each version gave it.

What we learned. The earlier model tracked supply and demand with no way for the market to respond. It matched 1995–2013 but predicted a surplus through the 2010s; in fact the market recovered by 2016 and was short by 2022.

History supports:

a market adjustment that is at least moderately fast and absorbs gaps of at least about 10% of demand;
tuned only on 1995–2013, it made the error on the test years less than half as large, mostly by pulling the market back toward balance.

History can't tell us:

how fast or large the adjustment could get, because past gaps never exceeded about 10–15%; the upper ends we use are judgments;
whether surpluses are really absorbed: that rests mostly on one recovery (2013–2018), and how work flows back to radiologists is assumed. If they are not, the 2045 oversupply chance is .

One caution: on the test years, our forecasts did worse than two naive guesses, a balanced market and no change from the last known state (there were only three test episodes). Those guesses can say nothing about AI, which is why we model it, but treat the long-range numbers as reasoned judgment, not a proven track record (full test).

What this changed

Version 1.5 put the 2045 oversupply chance at 42%. Three changes driven by this history bring it to :

recent imaging growth per person rises from 0.6% to 1.2% a year, between v1.5's claims-based estimate and the 1.8% the history implies, with a wider range, because wider ranges predicted past years better;
today's shortage is slightly milder, as reconstructed;
the market adjustment is added.

Removing the market adjustment while keeping the other two gives (the "no market adjustment" row under How sure?). That split is approximate, because the first two were themselves estimated with the adjustment on. So most of the drop from 42% comes from the market absorbing surpluses, the part of the adjustment history supports least.

How sure?

The real uncertainty is in our assumptions and model design, so the table re-runs the forecast under other defensible choices.

What holds and what doesn't. The direction holds in every row: oversupply risk is lower in 2035 than later and comes mostly from transformative AI. The exact numbers do not: the 2045 figure ranges from to , mostly because of assumptions about AI speed, imaging growth and the market adjustment. These ranges cover a chosen set of alternatives, not every plausible future.

The starting point is reconstructed, not measured. Today's balance (≈) combines market signals with the history reconstruction. With little or no shortage today, the 2045 oversupply chance would be (row above).

Backtest on other jobs. Run from 2016 for software developers, interpreters and translators, and medical transcriptionists, a simplified version of this method put every 2025 outcome inside its ranges, but a simple average of the official Bureau of Labor Statistics projection and the past trend was more accurate: our AI adjustment over-predicted job losses in medical transcription. (details; why a structured model)

What would change the forecast

If we observe X, how likely is meaningful oversupply? Some signals arrive within a few years, others later. Each row shows the chance of meaningful oversupply in the futures where that signal occurs; the small figure is how often the signal occurs.

These are signals, not levers. Early payment for AI-first reads mostly tells you that AI is moving fast, and fast AI is what raises the risk; on its own, faster regulation changes the 2045 figure much less (report §8.1).

Checkable predictions

Archived with this release (October 2026), with how each will be checked, so the forecast can be scored later.

Explore the simulated futures

Filter a random sample of 1,500 of the futures by what you believe, or re-weight the AI regimes. Everything updates instantly; numbers can differ from the rest of the page by a few points because of sampling.

AI-progress regime

Regulation, liability & payment

Fastest, middle and slowest third of futures by time from capability to payment for AI-first radiograph reads

Imaging growth per person

New AI-enabled imaging uses

Many new uses mostly occur in fast-AI futures (see signals, not levers)

Residency position growth

All filters except the regime split futures into thirds

Your regime weights

Defaults are the share of futures in each regime ( / / / ). Used when regime = Any.

Demand and supply in the matching futures

Median, 80% range, and 24 individual futures (thin lines)

Supply ÷ demand in the matching futures

All the numbers

The full forecast for reference.

Radiologist FTE demand and supply

Both indexed so 2026 demand = 1. Median line with 50% and 80% ranges; dashed line, demand with no further radiology AI. Hover for values.

Supply ÷ demand

Below 1 = shortage; above 1.10 = meaningful oversupply

Chance of oversupply or falling demand, by year

Share of the simulated futures

Median (10th–90th percentile). Demand and supply are indexed so 2026 demand = 1. The median of supply ÷ demand is not exactly median supply divided by median demand, because the two move together across futures.

How the model works

Five pieces are modeled separately and then linked, with a market adjustment that closes part of any gap between supply and demand. Each arrow carries uncertain inputs drawn from probability distributions, and the pieces are correlated: faster AI raises time savings and creates new imaging uses and speeds up scanners. Click a box for details.

Methods in brief

The "§" links above point here. The full report has every equation, table and source. We simulated futures and kept the not already contradicted by events (such as an FDA-authorized autonomous radiology read before now); with that many, simulation noise is under half a point.

Baseline demand

Radiologist work grows with demographics (+0.52%/yr, from projections adjusted for lower immigration), imaging per person (starting near %/yr and slowing, combining recent claims-based trends with the history reconstruction) and work per exam (0.4%/yr, slowing), minus up to 15% displaced by blood tests and other alternatives. Today's shortage (median , 80% range –) combines workforce projections and market signals with the history reconstruction. Report §4.1

AI time savings

Radiologist time is split into five tasks: interpretation 42%, drafting and measurement 18%, consultation 13%, administration 15%, procedures 12%. Each has an uncertain maximum AI time saving, a capability curve and an adoption curve. Maximums are anchored on measured effects: 15.5% faster documentation in live use, 42% faster reading in a reader study, and no significant pooled saving in real-world deployments. AI also adds oversight work. Adoption speeds up while there is a shortage. Report §4.2

Regulation and adoption

Each of the four tiers must pass validation, FDA authorization, and liability and payment acceptance before AI-first reading can begin, and an AI-first study still needs about a fifth of the radiologist's time for audit and sign-off (far less in the transformative-AI regime). Lags are anchored on MASAI (~5 years), the scarcity of clinical validation, autonomous retinal AI (3 years from FDA to payment), liability research, and the spread of computer-aided detection and electronic health records. Tier sizes come from autonomous-AI studies. Report §4.3

Jevons accounting

New work created by AI = cheaper reads (price elasticity ≈ −0.2 × fee share ≈ 20%) + faster turnaround + scanner throughput + new uses + follow-up of AI-found findings + new radiologist tasks − AI utilization management − reads shifting to other doctors, capped by scanner capacity. Time saved = baseline work × share of time AI saves. A Jevons paradox means new work exceeds time saved. Report §4.4

Supply

A cohort model tracks radiologists by years in practice (mean career ≈ 35 years; measured attrition 2.0–2.5%/yr in 2019–2022), calibrated so that flat residency positions reproduce the Neiman Institute's +25.7% by 2055. Positions (1,241 in 2026) grow about %/yr, a little below the 2010–2025 pace, and react to the market with a lag. New residents enter practice ~6 years after matching. Report §4.5

Market adjustment

Each year about of the remaining gap between supply and demand closes (80% range –) as work shifts between radiologists and other doctors and labor-saving tools spread faster or slower, up to a limit of about of demand (–). History bounds both only from below (fitted on 1995–2013, checked on 2015–2025); the upper ends are judgments (history test). Beyond the limit, a gap persists until residency positions respond. Report §4.7

Pay readout

Radiologist pay relative to other physicians changes each year in proportion to the imbalance (β ≈ – per unit of log supply ÷ demand; fitted to 2001–2014 and refitted on all years; it matches 2022–2025 only after we recoded the earlier pay data). Each year it also closes 5–20% of its gap to a normal level between today's and about 15% below it, before the recent shortage premium (both judgments). Imbalances beyond ±5% are treated as ±5%, since the fit saw only small ones. It does not feed back into supply or demand. Report §9.2

Uncertainty and regimes

uncertain inputs are drawn from explicit distributions and correlated through three shared factors: AI progress, regulatory friction and appetite for imaging. The AI factor picks one of the four regimes. Each input is graded anchored or subjective (see all). Report §4.6

Validation

The supply model reproduces published projections, unit tests check the accounting identities, the model is tested against radiology's job market since 1995 with a time split between fitting and held-out years, and a 2016→2025 backtest of three other jobs compares a simplified version of the method with official projections and trends. Report §6

Sensitivity analysis

Alternative assumptions and model designs (the table under How sure?) and, in the report, a tornado chart that pins groups of inputs at their 10th and 90th percentiles, variance shares and evidence-class attribution. Report §8

Every assumption, graded

Each input, its distribution and percentiles, and where it comes from. A anchored on data plus judgment, S subjective. Be most skeptical of the S rows: they account for most of the uncertainty.

References

AMA style, numbered in order of first appearance. Hover or tap a citation number in the text to see what the source supports and open it at the passage. "↩" links return to each place a source is cited.