What the study actually measured
The study, "Impact of Commercial Artificial Intelligence on Radiologist Reading Time for Pulmonary Nodule Evaluation at Chest CT," was published in RSNA's journal Radiology on September 1, 2026, by a team at Erasmus MC in Rotterdam, the Netherlands (Paramasamy et al.). It's a retrospective, single-institution, before-and-after analysis — not a randomized trial.
Researchers compared one full year of chest CT reads before the hospital deployed a commercial pulmonary-nodule detection and measurement tool against one full year after, covering September 2021 through May 2024: 19,433 patients and 39,323 exams in total (19,190 pre-AI, 20,133 post-AI). The primary outcome was "reporting time" — the interval from opening the exam in PACS to the radiologist authorizing the final report.
The result: adjusted median reporting time fell from 21.3 minutes to 18.2 minutes, a 14.6% reduction (p<.001), as first reported by AuntMinnie and Radiology Business.
"Up to 25%" is a subgroup, not the average
The 14.6% overall figure is the honest headline. The "up to 25%" framing comes from one subgroup, and the full breakdown shows the gains were far from uniform:
| Segment | Change in reporting time |
|---|---|
| All exams (overall median) | −14.6% |
| Thoracic radiologists | −25.0% |
| ECG-gated thoracic CT exams | −41.1% |
| Emergency department exams | +7.1% |
Two things stand out. First, the biggest gains clustered around readers and exam types where nodule work is a larger share of the read — thoracic subspecialists and gated thoracic protocols. Second, emergency department exams got slower, not faster, a reminder that a tool built for one workflow doesn't automatically speed up every workflow it touches.
What "reporting time" does — and doesn't — tell you
"Reporting time" here is a whole-workflow clock: PACS-open to report-authorized. That sounds like it should settle the "faster signed report" question. It doesn't, because of what the underlying tool actually does. The AI evaluated in this study detects, measures, and tracks pulmonary nodules on chest CT — it does not generate report text for the exam's other findings, and it doesn't draft the report itself. The radiologist still dictates or types the full chest CT report by hand; the AI's job is narrower: find the nodules, measure them, and flag change from a prior scan faster than a human scrolling through slices.
So the 14.6% reduction is real, but it's evidence that one time-consuming sub-task inside the read got faster — not evidence that a comprehensive, ready-to-sign report gets produced faster end to end. That distinction matters for anyone evaluating AI efficiency claims: "faster nodule call" and "faster signed report" are different claims, and this study, despite measuring a whole-workflow time metric, only directly supports the first.
It's also worth noting what the analysis didn't set out to do: it reported efficiency, not diagnostic accuracy or miss rate, and as a retrospective before/after comparison at a single center, it can't fully rule out other things that changed over the nearly three-year window — staffing, case mix, or workflow tweaks unrelated to the AI tool. None of that undermines the time-savings finding, but it does mean the study answers a narrower question than the "up to 25%" headline implies.
Not the same evidence as risk-scoring AI
It's also worth separating this from a different strand of nodule-AI evidence: studies that test whether AI can estimate a nodule's malignancy risk more accurately than tools like the Brock score. That's a classification-accuracy question, covered in a separate multicentre study xAID has covered before. This Erasmus MC study asks a workflow-speed question instead: given a nodule-detection tool a radiologist already trusts, how much faster does the read finish? The two questions — "is the AI's call more accurate" and "does the AI make the radiologist faster" — need separate evidence, and this study only speaks to the second.
Where this fits in the AI-CT-reporting landscape
Point tools like the one in this study are common precisely because they're easy to bolt onto an existing PACS workflow, and this data shows they can produce a real, if uneven, time benefit on the task they target. The alternative approach — and the one behind foundation-model CT reporting — is to draft the complete structured report across all findings in the study, not just nodules, with xAID's in-house radiologist reviewing every preliminary before it reaches the client's reading radiologist ready-to-sign. The efficiency case for that broader approach still needs its own head-to-head evidence; this study is a useful data point on what a narrow tool can and can't deliver, not a substitute for it.
Frequently asked questions
What did the new pulmonary nodule AI study actually find?
A retrospective study of 39,323 chest CT exams at Erasmus MC in Rotterdam, published in RSNA's journal Radiology, found that deploying a commercial pulmonary-nodule detection and measurement AI tool was followed by a drop in adjusted median reporting time from 21.3 to 18.2 minutes — a 14.6% reduction overall. The largest reduction, up to 25.0%, was seen among thoracic radiologists specifically.
Does the 'up to 25%' figure apply to every radiologist reading chest CTs?
No. The 25.0% figure was the largest subgroup reduction, seen in thoracic radiologists; ECG-gated thoracic exams saw an even larger 41.1% reduction. The overall average across all readers and exam types was 14.6%, and emergency department exams actually got 7.1% slower after AI deployment.
Did the study measure how much faster a full radiology report could be drafted?
Not directly. The study measured 'reporting time' as the interval from opening the exam in PACS to report authorization — a whole-workflow clock. But the AI tool itself only detects, measures, and tracks pulmonary nodules; it does not draft report text. Radiologists still wrote the full chest CT report themselves, so the time saved reflects a faster nodule sub-task inside an otherwise unchanged reporting workflow, not a faster-drafted, ready-to-sign report.
What does this mean for evaluating AI radiology efficiency claims?
It means efficiency headlines need context: a single-task detection tool produces single-task time savings that show up unevenly across exam types and reader subspecialties, and can even coincide with slower turnaround in some settings. Larger, more consistent gains generally require AI that drafts the complete structured report, ready-to-sign, not just one finding category within it.
Source: Paramasamy J, Odink AE, Mulders TA, et al. "Impact of Commercial Artificial Intelligence on Radiologist Reading Time for Pulmonary Nodule Evaluation at Chest CT." Radiology (2026). DOI: 10.1148/radiol.260484. As reported by Radiology Business and AuntMinnie. Figures are rounded as reported.