Abstract Early and accurate detection of intracranial aneurysms (IAs) is critical for preventing rupture; however, manual interpretation of time-of-flight magnetic resonance angiography (TOF-MRA) scans requires time-intensive review, increasing clinician workload. Although deep-learning methods offer promising solutions, existing approaches often rely on impractical vessel segmentation or inefficient patch sampling. Herein, we propose a dual-stream synergistic framework for IA detection and segmentation that balances diagnostic accuracy with clinical feasibility. The first stage performs landmark-guided candidate detection, beginning with a U-Net-based model trained using a novel adaptive loss function to predict the locations of 18 vascular landmarks. These landmarks guide the extraction of coordinate-aware patches, which are then analyzed using a hybrid UNETR-FPN model to identify potential aneurysm candidates. The second stream employs a fine-grained segmentation model (nnU-Net with a residual encoder) operating on the full MRA volume to delineate lesion boundaries. Final predictions are generated through conditional fusion of both streams, prioritizing candidate detection and refining shapes with segmentation information. The framework was trained and evaluated on 1055 TOF-MRA scans, including two distinct internal test sets. The landmark localization model achieved > 0.97 sensitivity in capturing the ground-truth aneurysms within generated patches. On the two test sets, the framework achieved lesion-wise sensitivities of 0.87 and 0.82 with corresponding false-positive rates of 1.23 and 1.17 per case. Performance was robust across anatomical locations but declined for aneurysms ≤ 3 mm. This landmark-guided dual-stream framework provides strong performance for IA detection while reducing annotation demands, offering a clinically practical tool for cerebrovascular diagnostics.
Kim et al. (Mon,) studied this question.