Deep learning assistance improves clinician performance in detecting cerebral aneurysms on MRA

Purpose or Learning Objective Reliable detection of intracranial aneurysms on MRA requires substantial expertise, yet diagnostic variability among physicians remains a challenge. As clinical use of MRA expands, there is a growing demand for consistent and accurate aneurysm detection. This study aimed to assess interobserver variability in detection performance across physicians with different levels of experience and to explore the potential role of artificial intelligence as a decision-support tool.
Methods or Background TOF-MRA studies with and without intracranial aneurysms were retrospectively collected. An inexperienced physician and a neurology resident with one year of training independently reviewed the images and marked suspected aneurysms. Their findings were compared with reference annotations provided by two board-certified neuroradiologists. After a washout period, the same readers reinterpreted the MRAs with the assistance of deep learning-based software. Diagnostic accuracy and interpretation time were compared between readings with and without AI support.
Results 74.5%. The neurology resident required 33 seconds per case, with a sensitivity of 60.0% and an accuracy of 66.8%. With AI support, the inexperienced physician improved to 42 seconds per case, with a sensitivity of 83.4% and an accuracy of 80.2%. The resident improved to 29 seconds per case, with a sensitivity of 85.9% and an accuracy of 86.1%.
Conclusion Deep learning-based AI assistance improved sensitivity and accuracy of intracranial aneurysm detection on MRA while reducing interpretation time.