A Landmark-Guided Dual-Stream Synergistic Framework for Automated Intracranial Aneurysm Detection in Magnetic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Early detection of intracranial aneurysms (IAs) is crucial to prevent rupture.
• Manual interpretation of time-of-flight magnetic resonance angiography (TOF-MRA) is time-consuming and increases clinician workload.
• Existing deep-learning methods for IA detection face challenges with impractical vessel segmentation or inefficient sampling.

Purpose of the Study:

• To develop and validate a dual-stream synergistic framework for accurate and clinically feasible IA detection and segmentation.
• To balance diagnostic performance with reduced annotation demands in cerebrovascular diagnostics.

Main Methods:

• A U-Net-based model predicts 18 vascular landmarks to guide coordinate-aware patch extraction.
• A hybrid UNETR-FPN model analyzes patches for potential aneurysm candidates.
• A fine-grained nnU-Net segmentation model delineates lesion boundaries on the full MRA volume.
• Conditional fusion of both streams generates final predictions, prioritizing detection and refining shapes.

Main Results:

• The landmark localization model achieved >0.97 sensitivity.
• The framework demonstrated lesion-wise sensitivities of 0.87 and 0.82 on two test sets, with low false-positive rates (1.23 and 1.17 per case).
• Performance was robust across anatomical locations but decreased for aneurysms ≤3 mm.

Conclusions:

• The proposed landmark-guided dual-stream framework offers strong performance for IA detection.
• This approach reduces annotation demands, presenting a clinically practical tool for cerebrovascular diagnostics.
• The framework aids in early and accurate detection, critical for preventing aneurysm rupture.