NCT07390708

Artificial Intelligence-Assisted Magnetic Resonance Imaging Diagnostic Strategy in a Tertiary Stroke Center

An Artificial Intelligence-Assisted Magnetic Resonance Imaging Diagnostic Strategy in a Tertiary Stroke Center-a Diagnostic Accuracy Study

• Status: Not Yet Recruiting
• Phase: —
• Study type: Observational
• Enrollment: 500 (estimated)
• Sponsor: Aarhus University Hospital · Academic / Other
• Sex: All
• Age: 18 Years
• Healthy volunteers: Not accepted

Summary

Quality improvement study with prospective observational design. The study monitors the diagnostic accuracy of an AI-assisted resident radiologist-termed the AI-ResRad diagnostic strategy-compared to an on-call specialist neuroradiologist-termed the SpecNeuroRad strategy-in interpreting stroke MRIs in patients with known onset. The study includes a pre-planned sub-study evaluating the diagnostic accuracy of neurologists and AI-assisted neurologists.

Detailed description

Current clinical practice and the supporting evidence base rely on interpretations by specialist neuroradiologists. Modern radiology departments face increasing imaging demands while contending with limited resources-including a shortage of specialist neuroradiologists. In the ideal setting, patients are evaluated in real time by a vascular neurologist and a neuroradiologist, who synchronously integrate clinical and imaging findings. In such cases, thrombolysis decisions can be re-evaluated concurrently with MRI acquisition, initiating treatment within minutes of scan completion. Although modern stroke MRI protocols can be completed in as little as 10 minutes, these rapid-response team activations often consume a disproportionate share of specialist time and availability. Consequently, real-world clinical practice frequently involves alternative team configurations, including resident radiologists, resident neurologists, and remote specialist consultations-compositions that vary depending on the on-call team's experience, time of day, and day of the week. Artificial intelligence (AI) can support the team with image interpretation, potentially optimizing time and resources. Recent studies have explored the role of AI-assisted stroke workflows and its ability to accurately detect ischemic lesions and hemorrhagic stroke-demonstrating promising encouraging diagnostic performance. However, there remains a need for prospective studies evaluating the real-world diagnostic accuracy of AI assistance as applied within its intended clinical use context To further understand the potential contributions of AI-assistance and resident radiologist interpretations, we designed the AID-STROKE accuracy study, under the Danish Quality Improvement legal and design framework. Sub-study: An Artificial Intelligence-Assisted Neurologist-based Diagnostic Strategy in Magnetic Resonance Imaging of Acute Stroke Patients with Known Onset-a Diagnostic Accuracy Study This pre-specified sub-study will be conducted in patients received at one of the hospitals (Gødstrup Regional Hospital)

Conditions

• Stroke
• Stroke Assessment

Interventions

Type	Name	Description
DIAGNOSTIC_TEST	Index test (AI-ResRad)	Eligible residents will review MRI sequences from the local PACS as they become available. During this process, they will maintain real-time communication-in person or via phone-with the treating neurologist, who will provide relevant clinical information. Simultaneously, the resident will have access to the AI output. Residents will have full access to the patient's electronic medical record and prior imaging. After integrating these inputs, the resident will complete an AI-assisted MRI interpretation using a predefined survey structure.
DIAGNOSTIC_TEST	Reference test (SpecNeuroRad)	Before any MRI sequences have been finalized, the resident will notify the on-call specialist neuroradiologist and pass on the clinical information received from the neurologist. The neuroradiologist will then independently review the MRI sequences as they become available in PACS, without access to the resident's interpretation or the AI results. They will also access the patient's electronic medical record and prior imaging. Once both parties have completed their respective surveys, the resident will call the neuroradiologist to jointly deliver an oral MRI interpretation to the neurologist. The radiologic information system's written radiology report may be completed by the neuroradiologist or the resident, with final sign-off by the neuroradiologist.
DIAGNOSTIC_TEST	Comparative Index tests: ResRad and AI-SpecNeuroRad	Two additional comparative test strategies will be evaluated: * ResRad: Prior to accessing AI results, residents will complete a non-assisted MRI interpretation. * AI-SpecNeuroRad: After completing their initial non-assisted interpretation (SpecNeuroRad), the specialist neuroradiologist will be granted access to the AI output. They will then submit a second interpretation (AI-assisted) using the same survey platform. To enforce internal blinding between non-assisted and AI-assisted interpretations, AI outputs are only revealed upon manual activation, which is timestamped by the system. Interpretation surveys are also timestamped at submission. All non-assisted interpretations must be submitted prior to AI output activation to be considered valid. Cases that violate this timestamp sequence will be excluded from final analyses to ensure methodological integrity.
DIAGNOSTIC_TEST	Neurologist Diagnostic Accuracy (Sub-study outcome)	Simultaneously with the radiologists' MRI interpretations with and without AI assistance according to the study workflow, the neurologist responsible for patient management will complete a similar MRI interpretation survey. The neurologist will first interpret the MRI without AI assistance and subsequently with AI assistance, mirroring the radiologist study workflow. The neurologist will be blinded to the radiologists' interpretations while completing their assessments, and likewise, the radiologists will be blinded to the neurologist's interpretations.

Timeline

• Start date: 2026-02-16
• Primary completion: 2027-07-01
• Completion: 2027-07-01
• First posted: 2026-02-05
• Last updated: 2026-02-09

Locations

2 sites across 1 country: Denmark