Multi-vendor AI technology at scale is tied to gains in turnaround times in high-volume imaging modalities, according to research published recently in the Journal of the American College of Radiology.
At about 23% to 25% annualized cost of one radiologist full-time equivalent (FTE) salary, multi-vendor AI generates 0.69 FTE of capacity through trauma radiography, according to researchers led by Sergey Morozov, MD, PhD, from Medlogic in Brussels, Belgium. But infrastructure latency is crucial.
“In about eight out of 10 exams, the AI result was already there before the radiologist even started the report,” Morozov told AuntMinnie. “And where the AI was well integrated, reports got done faster. For trauma x-rays, the median time went from five minutes to three.”
AI use is increasing in medical imaging, but this rise is uneven. In addition, many radiologists cite AI’s ability as a safety net to be valuable for subtle imaging findings, but many also report no major reductions in workloads.
Previous studies suggest that AI can reduce report turnaround times, albeit in narrow contexts. The researchers noted that the real-world impact of technical infrastructure on AI’s clinical utility is “poorly documented.”
“Everybody asks whether an AI algorithm is accurate. We wanted to know something more down to earth: does the answer actually reach the radiologist while they can still use it?” Morozov said. “Because if the AI result shows up after the report is signed, it’s not decision support anymore. It’s a very expensive screensaver, and a legal risk on top.”
Morozov and colleagues including Benoît Rizk from the 3R Swiss Imaging Network, investigated infrastructure latency, workflow, and radiologist sentiment across a 4.5-year, multi-vendor AI implementation program in a 20-center outpatient radiology network based in Switzerland.
The study included three retrospective cohorts: a technical cohort consisting of 96,874 exams performed between 2023 and 2025 for PACS-to-PACS latency and temporal alignment of 10 AI tools from seven vendors; a report turnaround time analysis cohort made up of 20,909 exams performed in 2025 comparing times between AI-available and concurrent non-AI workflows; and a survey cohort of 58 radiologists issued in 2025 in two waves.
The researchers reported an active AI adoption of 91.4% among the radiologists, with 35 (66%) saying that use AI regularly. They also found median total latency to be 2.06 minutes, with 72% of latency issues owed to data routing.
The “too late” rate, where the AI result arrived after reports are finalized, was 7.2% overall. This ranged from 3.0% for knee MRI to 13.2% for chest CT. After adjusting for radiologist activity, the researchers found AI availability to be linked to lower median turnaround times for trauma radiography and knee MRI (-26% and -18%; both p
Finally, in exploratory analysis, the net promoter score declined for chest CT (+38 to −3) and aorta CT (+22 to −25). Morozov said the results add to the literature focused on timing. He also offered advice to colleagues.
“You already have the timestamps in your systems. For each AI tool, check how often the result comes before, during and after the report,” he said. “It takes an afternoon, and it tells you whether you’re getting what you pay for.”
The researchers will next focus on whether radiologists use the result.
“We think two measures are the most practical. One is logging when a radiologist dismisses an AI result in the PACS, which the new IHE profile makes possible,” Morozov told AuntMinnie. “The other, once AI draft reports arrive, is how much of the AI text survives into the signed report. That’s the next step for us. We’re also working on continuous monitoring of delays with clear alert rules, and we plan to share our scripts openly once our partners have approved it.”
“So, the message is practical,” he continued. “We know where the problems are, we’ve measured them, and we know how to fix them.”
Read the full study here.