AI in Radiology: What the Evidence Actually Shows

Few areas of medicine have received as many AI products as radiology, and few have received as many inflated claims. For a hospital director or a ministry deciding where to invest, the useful question is not whether AI can read images. It is which uses have been tested in conditions resembling real practice, what those tests measured, and what they left unanswered. This article looks at three areas that are often proposed first for national programmes: breast screening, stroke triage and lung nodule detection.

A sober baseline: most products have little published evidence

A 2021 review in European Radiology mapped 100 CE-marked AI products for clinical radiology from 54 vendors. For 64 of them, the authors found no peer-reviewed evidence of efficacy. Where evidence existed, most studies measured diagnostic accuracy. Only 18 of the 100 products had evidence at the level of impact on diagnostic thinking, patient outcomes or costs. The market has grown since then, but the gap between regulatory clearance and proof of clinical benefit remains the starting point for any procurement.

A CE mark or an FDA clearance shows that a product meets safety and performance requirements for its stated purpose. It does not show that the product will improve outcomes in a particular hospital, on its own scanners, with its own patient population.

Breast screening: the MASAI randomised trial

The strongest evidence for AI in radiology comes from Sweden. The MASAI trial (Mammography Screening with Artificial Intelligence) randomised women in the national screening programme at four sites in southwest Sweden to either AI-supported screening or standard double reading by two radiologists. The AI system (Transpara 1.7.0) gave each examination a risk score. Low-risk examinations went to a single radiologist and the highest-risk ones to two radiologists, with suspicious areas highlighted. Radiologists made every recall decision.

The trial has now reported in three stages:

• Clinical safety analysis (Lancet Oncology, 2023), on 80,033 women: cancer detection was 6.1 per 1,000 with AI support and 5.1 per 1,000 with standard reading, with similar false-positive rates and a 44.3% lower screen-reading workload.
• Screening performance (Lancet Digital Health, 2025), on 105,934 women: detection rose to 6.4 versus 5.0 per 1,000 (a ratio of 1.29). The additional cancers were mainly small, lymph-node-negative invasive cancers, with no significant increase in recalls or false positives and a 44.2% reduction in screen readings.
• Primary endpoint (Lancet, 2026): the interval cancer rate, meaning cancers diagnosed between screening rounds that screening missed, was 1.55 versus 1.76 per 1,000. This met the trial's non-inferiority criterion. Sensitivity was higher with AI support (80.5% versus 73.8%), and specificity was the same in both groups (98.5%).

What MASAI does not tell you

MASAI tested one AI product in one screening programme that uses double reading, with experienced breast radiologists and a population-based invitation system. The finding that holds up best is the workload result: AI can safely replace one of two readers for most examinations. Programmes that already use single reading, have different equipment or serve different populations need their own validation. MASAI also did not measure mortality, which would require much longer follow-up.

Stroke: AI triage of large vessel occlusion

In ischaemic stroke caused by a large vessel occlusion (LVO), the benefit of thrombectomy falls quickly with time. AI tools that flag a suspected LVO on CT angiography and alert the stroke team by phone are designed to speed up the workflow. They are not designed to replace the radiologist's diagnosis.

The clearest test is a stepped-wedge cluster randomised trial at four comprehensive stroke centres in Houston, published in JAMA Neurology in 2023. Among 243 included patients with LVO stroke treated with thrombectomy, switching on automated LVO detection with secure group messaging was associated with an 11.2-minute reduction in door-to-groin time and a 9.8-minute reduction from CT to the start of treatment. There was no difference in thrombolysis times or length of stay. In an exploratory analysis, there was no significant difference in 90-day functional independence.

The intervention tested was software combined with a communication workflow, so the gains depend as much on alerting and team organisation as on the algorithm. The trial was also not powered to show better patient outcomes. For a health system, the business case for stroke AI therefore rests on a well-organised stroke pathway with a reachable thrombectomy team. Without one, an earlier alert has nowhere to go.

Lung nodules: strong accuracy data, thin outcome data

Low-dose CT screening of high-risk smokers has randomised evidence of benefit independent of AI. In the Dutch-Belgian NELSON trial (NEJM, 2020), screening reduced lung cancer mortality among men by 24% at ten years (rate ratio 0.76). Screening generates large volumes of CT scans and many small nodules, which makes it an obvious candidate for computer-aided detection (CADe) and nodule measurement.

The accuracy results are impressive. A widely cited deep-learning model published in Nature Medicine in 2019 reached an area under the curve of 94.4% on National Lung Screening Trial cases. When prior scans were not available, it outperformed six radiologists, with absolute reductions of 11% in false positives and 5% in false negatives. Results like this come from retrospective reader studies, however, not from prospective deployment.

A 2026 systematic review in European Radiology of commercially available AI for pulmonary nodule assessment on CT found 95 studies. The evidence base has matured: by 2024, more than a third of studies addressed higher levels such as diagnostic thinking, therapeutic decisions or patient outcomes. Yet all included studies had a high risk of bias in at least one domain, nearly two-thirds involved vendor funding or co-authorship, and evidence on patient outcomes and cost-effectiveness remains limited.

Five conditions for responsible deployment

In 2024 the radiology societies of the USA, Canada, Europe, Australia and New Zealand (ACR, CAR, ESR, RANZCR and RSNA) issued a joint statement on developing, purchasing, implementing and monitoring AI tools. Its themes, together with the studies above, suggest a practical checklist:

1. Local validation before go-live. Test the tool retrospectively on your own scanners, protocols and patient population, against radiologist reports and follow-up outcomes.
2. Radiologist-in-the-loop by design. Define exactly what the AI does, such as triage, flagging or measurement, and who signs off. In MASAI and the stroke trial, clinicians made every decision.
3. Workflow integration. Results must appear in the PACS and reporting system, and alerts must reach someone who can act on them. The stroke trial measured software plus messaging, not software alone.
4. Ongoing monitoring. Performance can drift with new scanners, protocol changes or shifts in case mix. Track agreement rates, false positives and missed findings continuously, not only at procurement.
5. Regulatory alignment. Under the EU AI Act, AI that is itself a medical device, or a safety component of one, and that requires third-party conformity assessment is classed as high-risk. Following the 2026 AI Omnibus amendments, those obligations apply from 2 August 2028. For an EU candidate country, procuring tools that will meet this standard avoids rework later.

What this means for Moldova

Moldova performs roughly 350,000–400,000 CT, MRI and PET-CT studies a year, and medical imaging is one of the five AI domains of the National Healthcare Digitalization Program (2025–2030). The evidence points to a sensible order of deployment. Begin with uses where AI has shown it can safely reduce workload or speed up time-critical pathways, run local validation on Moldovan data, and keep radiologists in charge of every report.

Two features of the programme help. First, because medical data must stay inside the country, both validation and production inference can run on dedicated GPU computing capacity hosted in Moldova. That keeps images and reports within national jurisdiction and lets performance be monitored against local outcomes. Second, a national programme can do what individual hospitals rarely can: build a shared, curated validation dataset and apply the same acceptance criteria to every tool before it reaches clinical use.

Related reading and solutions

• Medical imaging AI for CT, MRI and PET-CT
• Medical imaging equipment import
• Sovereign, on-premise AI infrastructure
• Sizing GPU infrastructure for healthcare AI
• Health data sovereignty: GDPR, EHDS and the EU AI Act

Planning an imaging AI pilot or a validation study? Contact Import Medical to discuss evidence-based deployment in Moldovan hospitals.

AI in radiology mammography screening stroke triage lung nodule detection clinical validation

Sources

1. Screening performance and characteristics of breast cancer detected in the MASAI trial (Lancet Digital Health, 2025) — The Lancet Digital Health
2. Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading in the MASAI study (Lancet, 2026) — The Lancet
3. Automated Large Vessel Occlusion Detection Software and Thrombectomy Treatment Times: A Cluster Randomized Clinical Trial — JAMA Neurology
4. Scientific evidence of commercial artificial intelligence products for pulmonary nodule assessment on CT scans: a systematic review — European Radiology
5. Artificial intelligence in radiology: 100 commercially available products and their scientific evidence — European Radiology
6. Developing, purchasing, implementing and monitoring AI tools in radiology: practical considerations. A multi-society statement from the ACR, CAR, ESR, RANZCR and RSNA — Journal of Medical Imaging and Radiation Oncology