The AI found twice as many nodules, took 15 seconds longer, and found no cancer

What happened

Radiology AI for lung nodules is usually sold with two promises in the same sentence: your radiologists will read faster, and they will miss less. The evidence behind both has come largely from retrospective reader studies, where radiologists sit down with a set of archived scans knowing they are in an experiment.

A team at Seoul National University Hospital, with a collaborator at National Jewish Health in Denver, tested the promises prospectively in the live reading workflow. Between May and September 2025 they randomised 911 consecutive people undergoing low-dose chest CT as part of a self-initiated general health checkup, 1:1, to interpretation with or without a commercial AI tool integrated into PACS.

The time saving did not happen. The extra detection did.

The test

Prospective, single-centre, parallel, open-label clinical trial. In the intervention group the commercial AI automatically detected, classified and measured nodules, and the results were displayed within PACS – the environment the radiologists use every day, not a research viewer.

Ten thoracic radiologists read the examinations. They reported only nodules measuring at least 4 millimetres, which is the standard reporting threshold for this population.

The primary outcome was interpretation time per examination. Secondary outcomes were the detection rates of Lung-RADS-positive nodules (categories 3 or 4), the detection rate of all nodules, and the frequency of recommendations for follow-up low-dose CT. Subsequent lung cancer diagnoses were recorded.

The numbers

Measure	AI-assisted	No AI	p
Individuals	447	464
Interpretation time per examination	187 s	172 s	.23
Lung-RADS-positive detection	16.9%	10.3%	.03
All-nodule detection	52.9%	32.6%	.002
Follow-up CT recommended	15.3%	7.4%	.04
Lung cancer diagnosed	0	0
Median follow-up	215 days	216 days

The result

Interpretation time was not significantly different, and the point estimate went the wrong way: 187 seconds with AI against 172 without.

Detection rose substantially. Lung-RADS-positive nodules were reported on 16.9 percent of AI-assisted scans against 10.3 percent without. All nodules rose from 32.6 to 52.9 percent.

Recommendations for a follow-up low-dose CT more than doubled, from 7.4 to 15.3 percent.

No individual in either group was diagnosed with lung cancer, over a median follow-up of 215 and 216 days respectively.

The authors conclude that the tool was not associated with a significant difference in interpretation times but was associated with significantly greater detection of clinically actionable nodules.

How this compares to what came before

Retrospective reader studies have reported both faster reading and better detection with nodule AI. The detection finding here is broadly consistent with that literature. The time finding is not.

The difference is almost certainly the setting. In a reader study, the radiologist’s task is to read the scans in front of them. In a live PACS workflow, an AI result is one more thing on the screen to evaluate, accept or override – and evaluating a machine’s suggestion takes time even when you agree with it. That cost does not show up when the comparison is retrospective.

This is the same pattern we saw in a large UK trial of chest X-ray AI prioritisation, where a feature that simulations predicted would save days saved nothing once tested in a live pathway. Live-workflow trials keep contradicting the retrospective ones, and in the same direction.

What it means

The business case for nodule AI needs rewriting. If it does not save reading time, the saving cannot be booked, and the value has to come entirely from the extra findings.

Whether the extra findings are worth having is genuinely unresolved by this trial. More Lung-RADS-positive nodules is the right direction. More than twice as many follow-up scans is a real cost – in radiation, in money, in the weeks of anxiety between being told there is something and being told it is nothing. And across 911 people followed for roughly seven months, there was no cancer in either arm to put on the other side of the ledger.

For screening programme design, the question this raises is about thresholds. A tool that surfaces every 4-millimetre nodule in an asymptomatic checkup population is working as specified. Whether that specification serves the population is a policy question that the detection rate alone cannot settle.

Business ideas from this paper

Follow-up burden modelling. Before deploying detection AI, a screening service should be able to forecast how many extra follow-up scans it will generate and what they will cost in capacity. This trial gives the multiplier: roughly double. Turning that into a planning tool is a straightforward product.

Live-workflow benchmarking. Reader studies and live trials now reliably disagree on time savings. A service that runs a short randomised live-workflow test inside a hospital’s own PACS, before purchase, would settle the question for that buyer.

Nodule follow-up management. If recommendations double, somebody has to track those patients, schedule the repeat scans and close the loop. That administrative layer is where the clinical risk sits and it is currently manual nearly everywhere.

Threshold configuration for screening populations. The same tool tuned differently for an asymptomatic checkup population versus a high-risk smoking cohort would produce very different follow-up burdens. Configuration against local population risk is consultancy work with a measurable output.

What could be wrong with it

No cancer was found in either arm, so the trial cannot say whether the extra nodules mattered. Median follow-up was about seven months, which is short for lung nodule surveillance.

Single centre, one commercial AI product, ten thoracic radiologists at a major academic hospital. Less experienced readers might show a different time profile, and plausibly a larger benefit from assistance.

The population was self-paying health-checkup attendees – healthier and lower risk than a screening programme targeting heavy smokers. Detection and follow-up rates will not transfer directly.

The study was open-label; radiologists knew whether they had AI, and knowing you are being timed changes how you read.

911 individuals is adequate for the time and detection outcomes and far too small for any cancer outcome.

How sure can you be?

We score this 8 out of 10.

It is a peer-reviewed prospective randomised controlled trial in the American Journal of Roentgenology, conducted in a live PACS workflow on consecutive real patients, with named institutions at Seoul National University Hospital and National Jewish Health and a registered trial identifier.

It loses points because it ran at a single centre with no independent replication, and the data are not publicly deposited.

The time and detection findings are well supported. Anything about whether patients benefited is not addressed by this trial at all.

Do this today

If you are evaluating nodule AI, ask the vendor for live-workflow timing data rather than reader-study timing data. If they only have the latter, this trial is the reason to discount it.

If you run a CT screening service, model your follow-up capacity at double the current recommendation rate before switching the tool on. That is the operational consequence this paper quantifies.

If you are a radiologist, the 187 against 172 seconds is worth knowing when someone tells you a tool will give you time back. In this trial, in the live workflow, it did not.

Source: Artificial Intelligence-Assisted Lung Nodule Evaluation on Low-Dose Chest CT in Asymptomatic Individuals: A Prospective Randomized Controlled Trial, American Journal of Roentgenology, March 2026. PMID 41848489 · doi.org ().

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.