Real-Time Pulsatile Flow Prediction for Realistic, Diverse Intracranial Aneurysm Morphologies using a Graph Transformer and Steady-Flow Data Augmentation

Extensive studies suggested that fluid mechanical markers of intracranial aneurysms (IAs) derived from Computational Fluid Dynamics (CFD) can indicate disease progression risks, but to date this has not been translated clinically. This is because CFD requires specialized expertise and is time-consuming and low throughput, making it difficult to support clinical trials. A deep learning model that maps IA morphology to biomechanical markers can address this, enabling physicians to obtain these markers in real time without performing CFD. Here, we show that a Graph Transformer model that incorporates temporal information, which is supervised by large CFD data, can accurately predict Wall Shear Stress (WSS) across the cardiac cycle from IA surface meshes. The model effectively captures the temporal variations of the WSS pattern, achieving a Structural Similarity Index (SSIM) of up to 0.981 and a maximum-based relative L2 error of 2.8%. Ablation studies and SOTA comparison confirmed its optimality. Further, as pulsatile CFD data is computationally expensive to generate and sample sizes are limited, we engaged a strategy of injecting a large amount of steady-state CFD data, which are extremely low-cost to generate, as augmentation. This approach enhances network performance substantially when pulsatile CFD data sample size is small. Our study provides a proof of concept that temporal sequences cardiovascular fluid mechanical parameters can be computed in real time using a deep learning model from the geometric mesh, and this is achievable even with small pulsatile CFD sample size. Our approach is likely applicable to other cardiovascular scenarios.

💡 Research Summary

Intracranial aneurysm (IA) rupture risk assessment increasingly relies on hemodynamic biomarkers such as wall shear stress (WSS) and oscillatory shear index (OSI), which are traditionally derived from computational fluid dynamics (CFD). CFD, however, is time‑consuming, requires specialist expertise, and is therefore unsuitable for routine clinical use. This paper introduces a deep‑learning surrogate that predicts the full cardiac‑cycle WSS vector field directly from the IA surface mesh, enabling real‑time inference. The core architecture is a Graph‑GPS transformer that combines a Graph Isomorphism Network with edge features (GIN‑E) for local message passing and a global self‑attention module for long‑range interactions. Geometry is encoded using Graph Harmonic Deformation (GHD) tokens, which provide a uniform, topology‑preserving down‑sampling of the mesh and serve as positional encodings. Temporal information is injected via a waveform encoder that processes the inlet flow profile and fuses it with the geometric features.

Training leverages the AneuG‑Flow dataset, comprising 14 000 steady‑state CFD cases (cheap to generate) and 808 pulsatile CFD cases (expensive). By augmenting the limited pulsatile data with the abundant steady‑state data, the model achieves a Structural Similarity Index (SSIM) of up to 0.981 and a maximum‑relative L2 error of 2.8 % on held‑out test aneurysms. Ablation studies demonstrate that (i) GHD‑based shape encoding markedly improves mesh fidelity and reduces computational overhead compared with naïve point‑cloud sampling, (ii) the hybrid local‑global attention is essential for capturing both fine‑scale wall‑shear variations and global inlet‑outlet effects, and (iii) steady‑flow augmentation substantially boosts performance when pulsatile samples are scarce (e.g., <100 cases).

Compared against prior methods—2‑D projection CNNs, point‑cloud GNNs, and earlier graph‑transformer models—the proposed framework consistently outperforms in quantitative metrics and visual similarity of predicted WSS fields. The authors also discuss practical deployment: once a patient‑specific mesh is reconstructed from imaging, fitting the GHD tokens takes only minutes, after which the network can generate the entire WSS time series in seconds on a standard GPU.

In summary, the study delivers (1) a scalable, mesh‑preserving deep‑learning pipeline for transient hemodynamic prediction in realistic IA morphologies, and (2) a novel data‑augmentation strategy that leverages inexpensive steady‑flow CFD to mitigate the scarcity of pulsatile training data. These contributions pave the way for integrating CFD‑derived biomarkers into clinical decision‑making for IA and potentially other cardiovascular pathologies.