{"id":19674,"date":"2026-10-01T05:09:05","date_gmt":"2026-10-01T08:09:05","guid":{"rendered":"https:\/\/rtmedical.com.br\/tmp-en-1790842145166\/"},"modified":"2026-10-01T05:09:11","modified_gmt":"2026-10-01T08:09:11","slug":"radiology-ai-real-world-latency","status":"publish","type":"post","link":"https:\/\/rtmedical.com.br\/en\/radiology-ai-real-world-latency\/","title":{"rendered":"Real-World Radiology AI: Faster Reports, Slow Pipes"},"content":{"rendered":"

Radiology AI in the real world<\/strong> cut report turnaround where volumes are high, but it ran into a bottleneck that has nothing to do with the algorithm: the network. An observational study in the Journal of the American College of Radiology<\/em> (JACR) followed Switzerland’s 3R Swiss Imaging Network, 20 outpatient centers running 10 AI tools from seven vendors, for 4.5 years. AI availability was associated with 26% faster reports for trauma radiography and 18% faster reports for knee MRI. In 7.2% of exams, however, the AI result arrived after the report had already been signed.<\/p>\n

\"Chart
Share of exams in which the AI result arrived after report sign-off, plus turnaround gains in the two highest-volume use cases.<\/figcaption><\/figure>\n

What the study measured, and how<\/h2>\n

The paper, by Sergey Morozov, Beno\u00eet Rizk and colleagues, describes a multi-vendor deployment run on the AI orchestration platform of France’s Incepto Medical. According to The Imaging Wire, roughly 389,000 AI-assisted exams went through the program between 2021 and 2025. For the analysis, the authors built three retrospective cohorts, each answering a different question.<\/p>\n

The technical cohort covered 96,874 exams from September 2023 to September 2025 and measured PACS-to-PACS latency: how long a study takes to leave the archive, get processed by the algorithm and come back as a result, and whether that result lands in time to be seen during the read. The turnaround time (TAT) cohort included 20,909 exams from March to August 2025 and compared AI-available workflows with concurrent non-AI workflows, first with a Mann-Whitney U test and then with linear mixed-effects models adjusted for radiologist. The third cohort was a two-wave 2025 survey of 58 radiologists using Likert scales and the Net Promoter Score (NPS).<\/p>\n

The headline numbers<\/h2>\n

After adjusting for radiologist, AI was associated with a 26% lower median TAT for trauma radiography<\/strong> and an 18% lower median TAT for knee MRI<\/strong>, both p<0.001. Brain volumetry MRI showed no significant change (+9.2%; p=0.33). On capacity, the program generated the equivalent of 0.69 full-time radiologists (FTE) in trauma radiography alone, a figure that falls to 0.46 FTE<\/strong> in the radiologist-adjusted sensitivity analysis. The authors put the program’s annualized cost at about 23% to 25% of one radiologist FTE salary.<\/p>\n

Uptake was high: 91.4% of radiologists (53 of 58) reported active use, and 66% of those used the tools regularly. Perceptions barely moved between waves, though. Per The Imaging Wire, the only shift was a modest gain in perceived productivity, from 2.57 to 2.94 on a five-point scale. In an exploratory analysis, NPS dropped for chest CT (from +38 to \u22123) and aorta CT (from +22 to \u221225), a result that was nominally significant only before correction for multiple comparisons.<\/p>\n

Latency: the bottleneck is routing, not inference<\/h2>\n

For anyone planning infrastructure, the most useful finding is the latency breakdown. Median total latency was 2.06 minutes per exam<\/strong> (interquartile range 1.74 to 3.05), and 72% of it<\/strong> was spent on data routing, meaning fetching the study from PACS, sending it to the AI service and returning the output, rather than on the inference itself. The “too late” rate was 7.2% overall, from 3.0% for knee MRI to 13.2% for chest CT, with 6.8% for trauma X-ray.<\/p>\n

As general technical context (not detailed in the abstract): in a typical setup the PACS or modality pushes the study over DICOM to an orchestrator, which dispatches it to the right algorithm, on premises or in the cloud, and receives back objects such as DICOM SR, secondary captures or annotations that must then be re-indexed in PACS and tied to the worklist. Every hop adds queuing, transfer and reconciliation. The authors’ conclusion that infrastructure latency, not algorithm speed, was the main barrier to clinical utility moves the conversation from accuracy to integration engineering. It is the same kind of hidden cost we covered when showing that PACS workstations draw power even when no one reads<\/a>: real-world performance depends on the whole ecosystem.<\/p>\n

There is a clinical reading too. Chest CT, with the highest late rate, is also where NPS fell the most. The study does not establish causation, but it is plausible that a result arriving after sign-off is perceived as noise, or even as extra work when it forces a signed exam to be reopened.<\/p>\n

Conflicts of interest and design limits<\/h2>\n

The paper discloses relevant conflicts, and they belong in any reading of it. Guerbet provided a research donation and Incepto supplied its Keros software free of charge; the 3R network holds royalties on the Keros knee MRI product distributed by Incepto. The primary finding involves Gleamer BoneView, used for trauma radiography, and a co-author holds equity in Gleamer. To mitigate bias, data extraction and statistical analysis were done by three authors with no financial ties to that company.<\/p>\n

The design is observational and retrospective, without randomization: comparing AI and non-AI workflows over the same period narrows, but does not remove, differences in case mix, shift or exam profile. The TAT window is six months, and the survey covered 58 radiologists at a single Swiss outpatient network. TAT gains are also not the same as clinical outcome gains, which the study did not measure.<\/p>\n

What it means for imaging groups<\/h2>\n

For multi-site outpatient groups with teleradiology and uneven network links, the lesson is straightforward: before comparing vendors on algorithm sensitivity, measure how long a study takes from modality to result back on the reading workstation, site by site. If the network already queues uploads to a central PACS, an excellent algorithm may simply arrive late. A per-modality “too late” rate is easy to compute by matching the AI result timestamp against report sign-off, and it belongs in every AI contract, including in emerging markets such as Brazil and the rest of Latin America, where connectivity between sites varies widely.<\/p>\n

The study supports the view that radiology AI needs a management plan first<\/a>: choose where volume justifies the spend, design the routing, and monitor use and delay after go-live. It also ties into the reporting rethink triggered by the PowerScribe 360 sunset<\/a>, since the moment an AI result appears in the dictation or viewer decides whether it is used at all. And in a market where radiology holds 76% of FDA AI authorizations<\/a>, real-world usage data like this is what separates regulatory clearance from operational value.<\/p>\n

Outlook<\/h2>\n

The 3R paper is one of few to report adoption, turnaround, FTE capacity and latency together across an entire multi-vendor network. Its key message is not that AI speeds up reports, which was already suspected for high-volume tasks such as fracture detection. It is that at scale, return depends on integration: up to one in eight chest CT results arrived when it no longer mattered. Shrinking that number is a job for IT, PACS and orchestration teams, not for data scientists.<\/p>\n

Source:<\/strong> The Imaging Wire<\/a> and Morozov et al., JACR 2026<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

JACR study at 20 Swiss centers: AI cut trauma X-ray report time 26%, but 7.2% of AI results arrived after the report was already signed.<\/p>\n","protected":false},"author":5,"featured_media":19635,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"ngg_post_thumbnail":0,"_rt_cluster":"","fifu_image_url":"","fifu_image_alt":"","footnotes":""},"categories":[102,101,100],"tags":[],"class_list":["post-19674","post","type-post","status-publish","format-standard","has-post-thumbnail","category-ai","category-pacs-en","category-radiology"],"aioseo_notices":[],"rt_seo":{"title":"","description":"JACR study at 20 Swiss centers: AI cut trauma X-ray report time 26%, but 7.2% of AI results arrived after the report was already signed.","canonical":"","og_image":"","robots":"index,follow","schema_type":"Article","include_in_llms":true,"llms_label":"Real-world radiology AI at Switzerland's 3R network (JACR 2026)","llms_summary":"Morozov et al., JACR 2026 (doi:10.1016\/j.jacr.2026.09.026): 4.5 years of multi-vendor AI (10 tools, 7 vendors, Incepto orchestration) at Switzerland's 20-center 3R outpatient network. Radiologist-adjusted TAT fell 26% for trauma X-ray and 18% for knee MRI; brain volumetry unchanged. 0.69 FTE capacity (0.46 adjusted) in trauma X-ray. Median latency 2.06 min, 72% routing; too-late rate 7.2% (3.0% knee MRI to 13.2% chest CT). Adoption 91.4% of 58 radiologists. Conflicts: Gleamer, Incepto, Guerbet.","faq_items":[],"video":[],"gtin":"","mpn":"","brand":"","aggregate_rating":[]},"_links":{"self":[{"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/posts\/19674\/"}],"collection":[{"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/posts\/"}],"about":[{"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/types\/post\/"}],"author":[{"embeddable":true,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/users\/5\/"}],"replies":[{"embeddable":true,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/comments\/?post=19674"}],"version-history":[{"count":1,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/posts\/19674\/revisions\/"}],"predecessor-version":[{"id":19679,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/posts\/19674\/revisions\/19679\/"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/media\/19635\/"}],"wp:attachment":[{"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/media\/?parent=19674"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/categories\/?post=19674"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/rtmedical.com.br\/en\/wp-json\/wp\/v2\/tags\/?post=19674"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}