Research page · Sergey Morozov, MD PhD MPH

Right result, right moment: monitoring that helps radiology AI deliver

What a 20-center network learned in 4.5 years, and five things any department can measure.

In a 20-center network, 91% of radiologists adopted AI, and where AI was integrated into the reading workflow, reports were finalized faster. The gap: 7.2% of AI results still arrived after the report was finalized, mostly because of data routing rather than the algorithm. Timing is something a department can measure every day and improve with routing and integration.

Watch: the study in 3.5 minutes

Narrated by Sergey Morozov. Independent explainer made by the authors from the published results; it contains no text, figures or tables from the journal article and is not produced or endorsed by JACR, the American College of Radiology or Elsevier. Transcript · Download (MP4, 21 MB)

What works

Morozov et al., Journal of the American College of Radiology, 2026. Multi-vendor AI program, 3R Swiss Imaging Network, 20 centers: ~389,000 AI-processed exams, 58 radiologists.

91%of radiologists adopted AI (active adoption 91.4%, 53 of 58)
-26% / -18%lower median report turnaround time where AI was integrated into the reading workflow: trauma radiography and knee MRI, adjusted for radiologist
0.69 FTEestimated annual reporting capacity from trauma radiography alone (0.46 FTE radiologist-adjusted), at an annual platform cost equivalent to 23-25% of one radiologist FTE salary
80.6%of AI results were available before the report was created

What needs work, and how to fix it

7.2% of AI results arrived after the report was finalized (the Too Late rate), while 80.6% were available before the report was created.What it means: these results came too late to change the report they belonged to. Action: measure the share of results available before reporting and the Too Late rate every month, and bring the late share down.
72% of the median 2.06-minute latency was data routing, not the algorithm.What it means: most of the delay sits in integration, so it can be reduced without changing the algorithm. Action: prefer automated push routing, and ask IT and each supplier for routing and inference time separately.
CT and MRI results arrived during or after report finalization more often than radiography results (29.6% vs 13.6%).What it means: in cross-sectional imaging, the largest share of results still arrives while the radiologist is reporting or after. Action: set time-to-display targets per modality in contracts and acceptance tests, and start routing work with CT and MRI.
Five of nine AI solutions had stable quarterly latency; four did not.The four showed progressive latency drift, post-deployment convergence, or a temporary infrastructure incident with full recovery. What it means: latency changes after go-live and is invisible without a routine check. Action: review latency per solution every quarter, keep metric definitions versioned so quarters stay comparable, and put latency targets in contracts.
Perceived productivity stayed below neutral (2.94 of 5) despite measured time savings.What it means: radiologists do not yet feel the gains that the timestamps show. Action: share the measured results with radiologists, and use training and regular feedback.

Observational study: associations, not causation.

Five actions for your department

  1. Measure the share of AI results available before reporting, and the Too Late rate. For every AI-processed exam, compare two HL7 timestamps: when the AI result became available and when the report was created and finalized.
  2. Set a time-to-display target in contracts and acceptance tests: the maximum minutes from end of scan to AI result on the reading screen, per modality.
  3. Prefer push routing and integration. Send studies to AI automatically at acquisition and return results into the tools radiologists already use. Routing, not the algorithm, is where most of the time goes.
  4. Review latency stability every quarter for each AI solution, with versioned metric definitions and radiologist feedback, and decide what to tune, upgrade or replace.
  5. Survey radiologists quarterly or every six months, not less often: trust, quality, productivity and a recommendation score for each tool, and share the measured results with them.

A monitoring set in three layers

The paper proposes monitoring deployed AI with metrics that come from data every department already has, plus a short survey:

1. Machine timeTotal latency from scan end to AI result in RIS/PACS, split into data routing and AI processing; quarterly median and coefficient of variation to check latency stability.
2. Workflow timeWhen the result arrives relative to the report: before creation, during dictation, or after finalization (Too Late rate); report turnaround time with and without AI.
3. Radiologist perceptionSurvey quarterly or every six months: trust, quality and perceived productivity (Likert), and Net Promoter Score per AI tool.

Reviewed quarterly in a Plan-Do-Check-Act cycle, this set supports the post-market monitoring that the EU AI Act requires for high-risk AI.

Measure it in your department

Open-source code too-late-rate (Python, Apache-2.0) computes latency, the arrival of each AI result relative to the report (including the Too Late rate) and a quarterly stability flag from HL7 timestamps. It runs on synthetic demo data out of the box and contains no patient data.

Read the article

Morozov S, Heracleous N, Korka D, Thouly C, Dufour B, Novarina O, Rizk B. AI Latency, Report Turnaround Time, and Adoption in a Multi-Vendor AI Ecosystem: A Multi-Site Observational Study. J Am Coll Radiol. 2026. doi:10.1016/j.jacr.2026.09.026

Article page of the paper on ScienceDirect, J Am Coll Radiol 2026
Screenshot of the article page on ScienceDirect (Elsevier), reproduced with permission.

Cite this study

Morozov S, Heracleous N, Korka D, Thouly C, Dufour B, Novarina O, Rizk B. AI Latency, Report Turnaround Time, and Adoption in a Multi-Vendor AI Ecosystem: A Multi-Site Observational Study. J Am Coll Radiol. 2026. doi:10.1016/j.jacr.2026.09.026

Download: BibTeX · RIS (EndNote, Zotero, Mendeley). Please cite the published article by its DOI.

Images for your talks and posts

Four ready-made summary images, redrawn from the authors' data. Each carries the citation, the DOI and a QR code to this page.

When the AI result arrivesSlide 16:9
Monitoring that helps AI deliver: three layersSlide 16:9
Four key numbersSquare, LinkedIn
Four key numbersPortrait, Instagram

All four as one PDF · Share this page on LinkedIn

For department and network leaders

A two-page brief: what worked, what needs work, and five things to measure in your own department. Executive brief (PDF)

Use of these materials

The accepted manuscript and the images are shared under CC BY-NC-ND 4.0. You may read, print and share them unchanged, show them in non-commercial teaching and talks, and post them on social media, always with credit (authors, title, journal, DOI, licence). You may not share changed versions or use them commercially, for example in vendor sales material, paid courses or marketing. For other uses of the article, request permission from the publisher (Elsevier). For other uses of the images or the brief, contact the author. The publisher's PDF is not covered by this licence; the published article is the version of record.

Talks on this topic

Upcoming

Past

Disclosures

Funding and conflicts of interest are declared in the article. S.M. provides R&D consulting services to 3R Swiss Imaging Network via Medlogic.