AI in Radiology: What the Trials Actually Show

How AI fits into a radiologist's day

In real programs, AI does one of three jobs. Triage: risk-scoring exams so low-risk scans get a single human read while suspicious ones get two. Detection support: highlighting suspicious regions for the radiologist as they read. Standalone reading: the AI reads alone — tested experimentally, not routine practice. Every deployed screening program keeps at least one human reader in the loop.

MASAI: the first randomized trial of AI mammography

MASAI is the landmark study — a randomized controlled trial of AI-supported breast-cancer screening in Sweden. The interim analysis (published in The Lancet Oncology, 2023) randomized 80,033 women and found AI-supported screening detected 6.1 cancers per 1,000 screened versus 5.1 with standard double reading, while cutting screen-reading workload by 44.3% — with false-positive rates unchanged at about 1.5% in both groups^[1].

The full results (The Lancet, January 2026, 105,915 women) went further. The key outcome: interval cancers — cancers diagnosed between screenings, the ones screening exists to catch — fell from 1.76 to 1.55 per 1,000 women, a 12% reduction in the AI arm. The AI group also saw 16% fewer invasive, 21% fewer large, and 27% fewer aggressive-subtype cancers at follow-up, with 81% of cancers detected at screening versus 74% in the control group^[2]. This is the strongest outcome evidence for AI in imaging to date — and the authors stress that AI-supported screening still requires at least one human radiologist, tested tools, and continuous monitoring^[2].

The supporting studies — and their caveats

In ScreenTrustCAD (Lancet Digital Health, 2023), one radiologist plus AI detected 261 cancers versus 250 for the standard two-radiologist read — a 4% improvement — while roughly halving radiologists' reading time, and AI alone was statistically non-inferior to two readers. The study was partly funded by the AI vendor (Lunit), and the authors still require a human radiologist before any recall^[3].

PRAIM (Nature Medicine, 2025) is the largest real-world study: 463,094 women in Germany's screening program, with radiologists choosing whether to use AI. AI-supported reading raised detection from 5.7 to 6.7 cancers per 1,000 (+17.6%) without meaningful recall increases. But it was observational — radiologists self-selected into using AI, so differences could come from who used it, not just the tool^[4].

Beyond mammograms: where AI is regulated and where it stumbled

Diabetic retinopathy holds a milestone: IDx-DR (now LumineticsCore) became, in April 2018, the first AI diagnostic authorized to make a screening decision without a clinician reading the image. In its registration study (900 patients across 10 primary-care sites), it correctly identified more-than-mild retinopathy 87.4% of the time and correctly cleared those without it 89.5% of the time — for one condition, on one specific camera, with a human workflow around it^[5]. Narrow, but real.

The cautionary tale is sepsis prediction. The widely deployed Epic Sepsis Model was evaluated externally (JAMA Internal Medicine, 2021) on 38,455 hospitalizations: it alerted on 18% of patients yet missed 67% of actual sepsis cases, adding false alarms that contribute to alert fatigue^[6]. Regulatory clearance and vendor marketing existed; independent validation told a different story. And for skin-cancer photo analysis, a Cochrane review concluded the evidence is limited and such systems cannot yet be relied on to rule out melanoma safely^[7].

Can AI diagnose cancer?

No. AI assists — it triages and flags. Diagnosis is a clinical decision made by a qualified physician, and no AI system is approved to diagnose cancer on its own. The MASAI authors were explicit: their results "do not support replacing healthcare professionals with AI"^[2]. What AI demonstrably does is shrink the repetitive part of screening — fewer normal scans per radiologist — while humans keep the judgment, the communication and the responsibility.

Frequently asked questions