Automated brain MRI protocol adaptation based on imaging findings using AI: diagnostic performance and agreement with
Kaining Sheng1,2, Alexander Cuculiza Henriksen1,2, David Mihal3
1Department of Radiology, Copenhagen University Hospital Rigshospitalet, Copenhagen, Denmark.
Acta Radiologica (Stockholm, Sweden : 1987)
|July 21, 2026
Summary
An AI tool shows promise in automatically adapting brain MRI protocols by detecting critical findings like infarcts and hemorrhages. While it demonstrates good pathology detection, expert oversight is still needed for optimal results.
Area of Science:
- Neurology
- Medical Imaging
- Artificial Intelligence in Medicine
Background:
- Selecting brain MRI protocols is often manual and error-prone.
- Automated AI analysis of initial sequences can enable dynamic protocol adaptation.
- Incomplete clinical data complicates manual protocol selection.
Purpose of the Study:
- To evaluate an AI tool's diagnostic performance in detecting critical brain findings.
- To assess the AI tool's agreement with neuroradiologists on protocol adaptation.
- To determine the AI tool's utility in adapting MRI protocols based on initial sequences.
Main Methods:
- Retrospective analysis of 752 brain MRI scans from two centers.
- AI tool and neuroradiologists independently assessed three initial sequences (DWI, T2-FLAIR, SWI/T2*-GRE).
- Comparison of AI recommendations against reference findings and neuroradiologist agreement (κ=0.47).
Main Results:
- AI demonstrated pooled sensitivity of 92% for infarcts, 75% for hemorrhages, and 71% for mass lesions.
- Pooled specificity ranged from 86% to 93% for the detected pathologies.
- High concordance (84%-87%) with neuroradiologists for scans needing no further adaptation.
Conclusions:
- The AI tool shows reasonable pathology detection and relevant protocol recommendations.
- It has potential for ensuring appropriate imaging protocols in high-volume, low-risk scenarios.
- Expert neuroradiologist oversight remains essential for optimal AI tool application.
