Graph and Geometric Deep Learning for Intracranial Aneurysm Geometry and Hemodynamics: A Systematic Review with...

Background: Neurosurgery depends heavily on implanted biomaterials, and the endovascular treatment of intracranial aneurysms (IAs) is the paradigmatic case, since coils, flow diverters and intrasaccular devices achieve durable occlusion only through intrasaccular thrombus organization and endothelial coverage of the neck, both governed by the local flow environment. Computational fluid dynamics (CFD) resolves that environment but is too slow and too solver-dependent for clinical or device design use. Graph and geometric deep learning operates natively on the unstructured meshes in which vascular anatomy and implanted scaffolds are represented and has been proposed as the technology that would close this gap. Methods: Four databases were searched up to 29 July 2026 for original studies developing or validating graph-based or geometric deep learning applied to IA geometry, with a hemodynamic or clinical outcome. Screening and extraction were performed in duplicate; risk of bias with PROBAST and artificial intelligence signaling items, reporting with TRIPOD+AI, and synthesis followed SWiM. Results: Twelve studies (2021 to 2026) were included: seven clinical (1965 aneurysms, plus 81 externally) and five computational (up to 984 geometries). Learned geometric representations discriminated rupture, growth and post-embolization recanalization better than morphological indices and then PHASES, reaching an area under the curve of 0.795 to 0.97 internally. Graph and point cloud surrogates reproduced hemodynamic fields with normalized errors of 2% to 5% in seconds rather than hours, and one transformer-based graph network predicted the extent and timing of intra-aneurysmal thrombus formation over a reactive surface. External validation was reached by two studies, and clinical utility was achieved in one; discrimination fell from 0.85 to 0.71 in one external test, and surrogate error rose from 4.1% to 19.1% on patient-derived anatomy. No model was trained on a device-laden geometry. Additionally, 5/7 clinical studies classified cross-sectional rupture status rather than prospectively predicting future rupture. Conclusions: Graph and geometric deep learning has substantially reduced the computational burden of hemodynamic analysis, but reliable generalization to unseen patient-specific anatomy remains a relevant challenge. Shared patient-specific benchmarks, prespecified external validation, reporting of calibration, and extension of training corpora to implanted anatomy are necessary next steps towards virtual evaluation of neurovascular biomaterials.