- Design
- multi-site retrospective observational study across three cohorts, with linear mixed-effects adjustment for radiologist
- Population
- 96,874 examinations for latency, 20,909 for turnaround time, 58 radiologists surveyed, across 20 centres
- Primary outcome
- archive-to-archive latency, report turnaround time, adoption and sentiment
- Effect
- median latency 2.06 min (72% routing); results too late in 7.2% overall and 13.2% for chest CT; turnaround -26% trauma radiography, -18% knee MRI
A 20-centre outpatient radiology network ran 10 artificial intelligence tools from 7 vendors over four and a half years. Three cohorts were analysed: 96,874 examinations for infrastructure latency, 20,909 for report turnaround time against concurrent non-artificial-intelligence workflows, and 58 radiologists surveyed twice.
Median time from one picture archiving system to the other was 2.06 minutes (IQR 1.74-3.05), and 72% of that was data routing rather than computation. The consequence has a name in this study: the result arrived after the report was finalised in 7.2% of cases overall, ranging from 3.0% for knee MRI to 13.2% for chest CT. One chest CT in eight had its artificial intelligence output land too late to be used.
Where the timing worked, the gains were real. After adjusting for individual radiologist, turnaround time was 26% lower for trauma radiography and 18% lower for knee MRI (both P < 0.001); brain volumetry MRI showed no change. The economics are stated plainly: at around a quarter of one radiologist's annual salary, the programme generated 0.69 full-time equivalents of capacity through trauma radiography alone, or 0.46 on a more conservative analysis.
Radiologist sentiment went the other way where the tools were slowest. Net Promoter Score fell for chest CT (+38 to -3) and aorta CT (+22 to -25), which the authors flag as exploratory and nominally significant before correction for multiple comparisons. Adoption remained high overall — 91.4% active, 66% of those using the tools regularly.
The practical lesson for any department evaluating procurement is that the vendor comparison most departments run is the wrong one. Algorithm performance was not what determined clinical utility here; routing was.
- Measure result latency end to end, from archive to archive, rather than accepting a vendor's inference time
- Track the proportion of results arriving after report finalisation — it was 13.2% for chest CT here
- Target routing and integration before algorithm selection; 72% of latency was data transfer
- Expect turnaround gains in high-volume, fast-reported studies such as trauma radiography, not in every modality
- Watch radiologist sentiment as an implementation metric; it fell for the modalities where results arrived late
Why it matters
Departments choose these tools on published accuracy, and accuracy was not what decided whether they could be used.
Don't overread it
The capacity figures come from one outpatient network's case mix and cost base, and do not transfer to a different setting.
The statistics, in plain English
This is an observational comparison of workflows that were available or not available, not a randomised allocation, so cases reported with the tool available may differ from those reported without it; adjustment for individual radiologist removes one source of that but not all. The turnaround time comparison used a rank-based test appropriate for skewed timing data. The sentiment findings are explicitly exploratory and were only nominally significant before correcting for multiple comparisons, which means they should be treated as a signal to investigate rather than a result.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for radiology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free