Fully Automated Deep-Learning and CFD Framework for intracranial Aneurysm hemodynamic Analysis

Introduction: : An intracranial aneurysm (IA) is a pathological dilation of a cerebral artery resulted from localized weakening of the vessel media, which increases the risk of rupture and catastrophic subarachnoid hemorrhage (SAH). In the United States alone, aneurysmal SAH affects approximately 30,000 patients annually, with a mortality rate nearing 50% and significant long-term neurological impairment among survivors. Patient-specific computational fluid dynamics (CFD) has emerged as an indispensable tool for detailed quantification of intra-aneurysmal flow patterns and wall shear stresses (WSS), offering valuable insights for risk assessment and treatment planning. Despite its clinical potential, conventional CFD workflows remain labor-intensive, technically demanding, and time-consuming, limiting their adoption in acute care settings. To overcome these limitations, we have developed a fully automated, end‐to‐end pipeline that integrates open‐source tools for segmentation, meshing, simulation, and post‐processing into a single, clinically adaptable platform.

Materials and

Methods: : We begin by applying a deep learning-based segmentation network to generate a coarse mask of the aneurysm and surrounding vessels. This mask is then refined through a two-stage post-processing pipeline: Stage 1 employs an nnU-Net architecture as a primary landmark detector to localize the aneurysm region within the vascular tree. The resulting output is used to clip adjacent vessels, reducing geometric complexity and focusing the region of interest. Stage 2 uses PointNet++ to further refine the aneurysm geometry, enabling precise lesion isolation (Fig.1). This delineated aneurysm domain is then used for morphological characterization and defining the domain for hemodynamic analysis. Moreover, the centroid of the isolated aneurysm serves as a landmark to truncate downstream vessels. Following these steps and additional post-processing including surface smoothing and flow extensions, high-quality volumetric meshes are generated using the open-source TetGen package. Hemodynamic simulations are conducted by assigning appropriate blood rheology and accurate boundary conditions, and executed on our high-performance computing cluster using GPU-accelerated PyFluent.. Computational results are post-processed with custom Python scripts built on the Visualization Toolkit (VTK) framework. To validate robustness and accuracy, the entire automated workflow is replicated manually to create ground-truth data and comparative statistical analyses confirm the fidelity of the automated pipeline relative to manually-curated benchmarks.

Results, Conclusions, and Discussions:: Preliminary validation involved quantitative comparisons between morphological metrics derived from manual and deep learning–based segmentations. The automated approach demonstrated robust accuracy, achieving Pearson correlation coefficients exceeding 0.95 with minimal relative error across all evaluated geometric indices (Fig.2). Further validation of hemodynamic outputs, specifically velocity fields and WSS, yielded intraclass correlation coefficients ranging from 0.85 and 0.95, reflecting good to excellent concordance with the manual reference standard. Notably, the simulated WSS and velocity streamlines exhibited near-identical spatial patterns between manual and automated workflows (Fig.3). Collectively, these findings confirm that the automated pipeline faithfully reproduces conventional CFD results while substantially reducing the requirement for expert user input.
By eliminating operator variability in segmentation, meshing, and preprocessing, our automated pipeline ensures reproducibility and consistency in hemodynamic results regardless of user expertise. External validation using independent datasets confirmed the pipeline’s generalizability, with deviations from manual benchmarks well within acceptable thresholds. The reduction in manual labor significantly accelerates workflow throughput and enables large-cohort analyses, while seamless integration of deep-learning and VTK tools supports scalable deployment in both research and clinical settings. Ongoing development includes real-time visualization dashboards and expansion to other cerebrovascular disorders.

Acknowledgements and/or References (Optional):: The fully automated pipeline demonstrated excellent agreement with conventional CFD workflows for cerebral aneurysms, achieving Pearson correlation coefficients > 0.95 for geometric metrics and intraclass correlation coefficients of 0.85–0.95 for hemodynamic outputs. Leveraging GPU-accelerated PyFluent simulations, total runtime was reduced by approximately one order of magnitude compared to CPU-only workflows lowering per-case processing from several hours to just minutes. This marked performance gain facilitates high-throughput, patient-specific evaluation of large cohorts and brings CFD-based risk assessment closer to near-real-time clinical decision support. Future extensions will integrate interactive visualization platforms and broaden the pipeline’s translational impact.