AttendMe Owl Logo
AttendMe
Evidence Evolution
RadiologyRadiology

How This Evidence Evolved

AI in Diagnostic Radiology

From computer-aided detection to deep learning and the augmentation versus automation debate

19.4

Timeline

Signal

Early observations and pilot data that first suggested a new direction

AI in radiology needs evaluation at several levels: diagnostic accuracy, workflow and patient outcomes. A review described potential roles in reading efficiency, disease detection and personalized diagnostics, while concluding that little was known about its contribution to clinical practice. Traditional mammography CAD also requires its own evidence: a result should not be attributed to DMIST without checking the actual publication.
Proof

Landmark RCTs and pivotal trials that established the evidence base

Of 903 AI-enabled medical devices listed by the FDA as of August 31, 2024, most became available in the last decade, and 692 (76.6%) were in radiology. Clinical performance studies at approval were reported for 505 devices (55.9%); of these, only 41 (8.1%) were prospective and 12 (2.4%) randomised. Evidence of clinical impact differed by use case. Among 243 eligible patients in a stepped-wedge cluster randomised trial, automated large vessel occlusion (LVO) detection coupled with secure messaging was associated with an 11.2-minute reduction in door-to-groin time (95% CI -18.22 to -4.2). A prospective evaluation of AI triage for intracranial hemorrhage (9,954 examinations) found no improvement in report turnaround time for ICH-positive examinations (147.1 minutes without AI vs 149.9 with AI; p = .11).
Extension

Follow-up studies, subgroup analyses, and real-world validation

MASAI tested AI-supported mammography reading using Transpara version 1.7.0. Risk scores 1–9 received single reading; score 10 retained double reading. In its 2023 safety analysis, cancer detection was 6.1 versus 5.1 per 1,000 with standard double reading (ratio 1.2; 95% CI 1.0–1.5; p=0.052), with 44.3% less screen-reading workload. Interval cancer follow-up remained pending in that report. This protocol did not replace one reader for every examination. A separate chest-radiograph study found selective underdiagnosis in underserved populations across three large datasets and a multisource dataset, including higher rates in intersectional groups. These findings support evaluating workflow benefits and demographic performance separately.
Now

Current standard of care and ongoing research directions

Clinical effects vary by task and setting. In the cluster randomized LVO study, AI detection with secure messaging was associated with an 11.2-minute reduction in door-to-groin time (95% CI, −18.22 to −4.2). In the prospective ICH study, mean report turnaround time was 147.1 minutes without AI versus 149.9 with AI (P=.11). A review called for real-world monitoring to inform procurement and reimbursement. A study of FDA-listed devices as of August 2024 found clinical performance studies for 505 devices (55.9%); these public summaries did not establish generalizability for every device. How should local validation, workflow effects and patient outcomes guide deployment?

Landmark Trials in This Story

Pediatric radiology 2022

How does artificial intelligence in radiology improve efficiency and health outcomes?

Narrative review: little known of AI's clinical contribution

Pediatric radiologyNarrative review
LVO software cluster RCT2023

Automated Large Vessel Occlusion Detection Software and Thrombectomy Treatment Times: A Cluster Randomized Clinical Trial.

Door-to-groin time: with vs without automated LVO detection software, 11.2 minutes shorter (95% CI -18.22 to -4.2)

JAMA NeurolStepped-wedge cluster randomised trialN=243
Prospective ICH AI triage2024

Prospective Evaluation of Artificial Intelligence Triage of Intracranial Hemorrhage on Noncontrast Head CT Examinations.

Report turnaround time for ICH-positive examinations: with AI 149.9 vs without AI 147.1 minutes (p = .11)

AJR Am J RoentgenolProspective two-phase evaluation (without AI, then with AI); 9,954 examinationsN=7,371
FDA AI device analysis2025

Generalizability of FDA-Approved AI-Enabled Medical Devices for Clinical Use.

Clinical performance studies at approval: reported for 505 of 903 devices (55.9%); 41 (8.1%) prospective, 12 (2.4%) randomised

JAMA Netw OpenCross-sectional study of FDA-listed AI-enabled devices (as of 31 August 2024)N=903
MASAI2023Landmark

Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study

Interval cancer rate (pending); detection 6.1 vs 5.1 per 1000 (ratio 1.2, 95% CI 1.0-1.5; p=0.052)

Lancet OncolRCTN=80,033
Nature Medicine 2021

Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations

Underdiagnosis rates by demographic group for state-of-the-art chest X-ray classifiers (4 datasets)

Nature MedicineAI dataset evaluation

Explore the evidence yourself

Ask AttendMe about any trial, guideline, or clinical question. Evidence-ranked answers from 3.7M+ peer-reviewed articles.

Related Evidence

Frequently Asked Questions

How many AI products are FDA-cleared for radiology?+
A review of FDA-listed AI-enabled devices through August 31, 2024 analyzed 903 devices, including 692 in radiology (76.6%). Clinical performance studies were reported for 505 devices (55.9%); the public information often did not permit comprehensive assessment of generalizability. Regulatory listing alone does not establish improved patient outcomes in a particular clinical setting.
Has AI in radiology demonstrated improved patient outcomes?+
Randomised evidence covers stroke triage and breast screening. In a cluster randomised trial, automated large vessel occlusion detection software reduced door-to-groin time by 11.2 minutes (95% CI -18.22 to -4.2). For intracranial hemorrhage, a prospective evaluation found AI triage did not improve report turnaround time (147.1 vs 149.9 minutes; p = .11). In MASAI, AI-supported screening detected 6.1 vs 5.1 cancers per 1000 with double reading, with 44.3% less screen-reading workload. However, large-scale evidence of improved survival or reduced morbidity from AI in radiology remains limited.
Will AI replace radiologists?+
The supplied studies evaluate specific forms of radiologist support, not replacement of comprehensive practice. MASAI reduced reading workload using selective single reading while retaining double reading for the highest AI risk score. Stroke software shortened treatment workflows, whereas a prospective intracranial-hemorrhage study found no improvement in diagnostic performance or reporting time. Clinical teams therefore need to assess each tool’s task, workflow, demographic performance and patient outcomes rather than assume a universal benefit.

Medical Disclaimer: This content is for educational purposes only and does not constitute medical advice. Clinical decisions should always be based on individual patient assessment, local guidelines, and professional judgement.

All data sourced from published, peer-reviewed articles and clinical practice guidelines.

Claims checked against cited sources: 10 November 2026