Please enable JavaScript.
Coggle requires JavaScript to display documents.
Radiology AI—Evidence and Adoption - Coggle Diagram
Radiology AI—Evidence and Adoption
⚡ Benefits
Improve diagnostic accuracy
Streamline workflow
📊 Adoption status
Numerous AI solutions exist
But widespread adoption is limited
Antonissen et al., 2025
639 supporting papers (↑ from 237)
Most of the studies have level 2 evidence only
#
173 CE-certified AI products (↑ from 100 in 2020)
Few studies at Levels 3–6
❓ Critical concerns
Uncertainty about clinical value
🛑 Barriers
Unclear ROI
Workflow integration
Data privacy
Framework To Assess The Efficacy of Diagnostic Imaging (Fryback & Thornbury, 1991)
Six levels of efficacy
Level 3: Diagnostic thinking– impact on clinician’s diagnostic decision
Level 4: Therapeutic impact– influence on treatment planning
Level 5: Patient outcomes– impact on patient health outcomes
Level 1: Technical– image quality, technical accuracy
Level 6: Societal impact– cost-effectiveness, accessibility, societal benefit
Our study
Level 2: Diagnostic accuracy & generalisability
Retrospective (2 hospitals)) → Evidence for diagnostic accuracy
This retrospective study trained and validated the AI using MRI data from two hospitals, achieving high diagnostic accuracy for urgent abnormalities and showing potential for clinical triage — corresponding to Level 2 in Fryback’s model
AI for Head MRI: Retrospective Study (
MIA
)
Problem Context
Global shortage of radiologists
Increased reporting delays
Growing demand for head MRI
Consequences: poorer patient outcomes, inflated healthcare costs
Proposed Solution
Role: flag abnormalities at time of imaging
Aim: prioritise reporting of abnormal scans
Computer vision models (AI)
Challenges in Development
Difficulty obtaining large, clinically-representative labelled datasets
Bottleneck for AI progress
Study Contribution
Input: minimally processed T2-weighted & DWI scans
Training: Transformer-based report classifier to auto-label data
Developed deep learning framework (CNN-based)
Dataset
Size: 70,206 examinations
Source: two large UK hospital networks
Results
Accuracy: AUC > 0.9
Generalisability: stable performance between hospitals (ΔAUC ≤ 0.02)
Speed: classification in < 5s
Interpretability: model outputs explainable
Impact Simulation
Reduction in mean reporting delay for abnormal scans:
2 more items...
Prospective (44 hospitals) → Evidence for generalisability
Cluster RCT (vendor-neutral) → Levels 3–66
Level 4: Therapeutic impact (impacting treatment decisions via reprioritization)
Level 5: Patient outcomes (faster diagnosis, improved management)
Level 3: Diagnostic thinking (AI reprioritizing brain MRI scans)
Level 6: Societal impact (equity of access, scalable across NHS)
Level 2: Diagnostic accuracy
#
– correctness of disease detection
Medical Device Evaluation Lifecycle
Medical Device Evaluation Lifecycle
Intended Use
→ What is the AI supposed to do?
Target population
Clinical pathway
Intended Medical Indication
Intended User—who is intended for
Device Function—what the device does
Scientific Validity
→ Is there evidence linking the AI output to the clinical condition?
Literature Review
Proof of concept studies
Analytical Validity
→ Does the technology work technically as intended?
Testing in controlled environment
Clinical Validity
→ Does it work in real patients and real NHS settings?
Clinical Investigation
Testing on humans
Target population
Real world settings
Regulatory Approval
→ Has it met UK regulatory requirements?
Post-Market Surveillance
→ How will safety and performance continue to be monitored?
Monitoring safety
Adverse events
Monitoring performance
Product lifetime
Essential Documentation
Clinical Investigation Plan (CIP)
Clinical Evaluation Report (CER)
Clinical Evaluation Plan (CEP)
Post-Market Surveillance (PMS) Plan:
Post-Market Clinical Follow-up (PMCF) Plan