Abstract 416: Automatic Segmentation of Intracranial Aneurysms: A Systematic Review of Methodologies and Performance Benchmarks.

Introduction Accurate segmentation of intracranial aneurysms (IAs) is a prerequisite for reliable morphometric analysis, rupture risk assessment, and computational fluid dynamics (CFD) modeling. Manual annotation, though considered the reference standard, is labor‐intensive and prone to inter‐observer variability. Automatic segmentation using deep learning (DL) and radiomics‐based algorithms has emerged as a promising approach. This systematic review evaluates current methodologies for automatic IA segmentation and benchmarks their performance across internal and external validation settings. Methods We systematically reviewed multicenter studies, retrospective validations, and prospective trials reporting automatic segmentation of IAs on CT angiography (CTA), MR angiography (MRA), and digital subtraction angiography (DSA). Extracted performance metrics included Dice similarity coefficient (DSC), sensitivity, and computational efficiency. Comparative analysis was performed across convolutional neural networks (CNNs), U‐Net architectures, vessel‐attention models, and hybrid approaches. Internal versus external validation outcomes were synthesized to assess robustness and generalizability. Results Segmentation methodologies were dominated by U‐Net variants and 3D CNN architectures. You et al. (2024) introduced a vessel‐attention U‐Net (VA‐Unet), achieving an internal Dice coefficient of 0.78 and external Dice of 0.71, with sensitivity of 0.962 for aneurysm detection. Wei et al. (2024) demonstrated comparable accuracy, reporting internal DSC of 0.75 and external DSC of 0.70 across multiple centers, with mean processing time of 1.7 minutes per scan. Studies using hybrid radiomics‐DL pipelines showed improved delineation of small and irregular aneurysms, with DSC ranging 0.72‐0.76. Overall, pooled analysis across available reports yielded a mean internal DSC of 0.76 (95% CI: 0.73‐0.79) and mean external DSC of 0.71 (95% CI: 0.67‐0.74), reflecting a performance drop of ∼0.05 on external validation. Sensitivity consistently exceeded 0.90 across models, but false positives were more frequent in distal or stenotic vessels. Computational efficiency varied: lightweight CNNs processed volumes within 30‐60 seconds, whereas more complex attention‐based models required >1 minute per case. Prospective studies confirmed feasibility, though segmentation accuracy was affected by image quality variability and aneurysm size (<3 mm lesions remained challenging). Conclusion Automatic segmentation of IAs using DL demonstrates strong accuracy, with Dice coefficients consistently above 0.70 and sensitivity exceeding 90%. U‐Net and attention‐based architectures represent the current benchmarks, though performance declines modestly on external datasets. These models are sufficiently accurate for clinical support and CFD integration, yet challenges remain in small aneurysm delineation and generalizability. Future work should emphasize harmonized multi‐center datasets, lightweight architectures for clinical deployment, and integration with rupture risk modeling to fully exploit segmentation as a clinical decision‐making tool.