Graph Deep Learning for Intracranial Aneurysm Blood Flow Simulation and Risk Assessment

TL;DR

The paper introduces a 51-million-parameter graph-transformer GNN that autoregressively predicts full-field cerebral hemodynamics from patient-specific vascular geometries, achieving CFD-like accuracy for velocity, WSS, and OSI on both synthetic and clinical aneurysm datasets while delivering an approximate 200-fold speed-up. Trained on AnXplore and few-shot geometries and validated on MATCH cases, the method generalizes to out-of-distribution anatomies and inflow conditions, enabling near real-time, image-guided risk assessment and potential integration with clinical workflows. The study also demonstrates meaningful alignment between GNN-derived hemodynamic risk metrics and CFD ground truth, as well as complementarity to the PHASES clinical score, suggesting a path toward rapid, interpretable decision support for aneurysm management. Future work includes physics-informed losses, boundary-aware learning, and scaling to larger meshes and treatment scenarios (e.g., stents, FSI) to broaden clinical impact.

Abstract

Intracranial aneurysms remain a major cause of neurological morbidity and mortality worldwide, where rupture risk is tightly coupled to local hemodynamics particularly wall shear stress and oscillatory shear index. Conventional computational fluid dynamics simulations provide accurate insights but are prohibitively slow and require specialized expertise. Clinical imaging alternatives such as 4D Flow MRI offer direct in-vivo measurements, yet their spatial resolution remains insufficient to capture the fine-scale shear patterns that drive endothelial remodeling and rupture risk while being extremely impractical and expensive. We present a graph neural network surrogate model that bridges this gap by reproducing full-field hemodynamics directly from vascular geometries in less than one minute per cardiac cycle. Trained on a comprehensive dataset of high-fidelity simulations of patient-specific aneurysms, our architecture combines graph transformers with autoregressive predictions to accurately simulate blood flow, wall shear stress, and oscillatory shear index. The model generalizes across unseen patient geometries and inflow conditions without mesh-specific calibration. Beyond accelerating simulation, our framework establishes the foundation for clinically interpretable hemodynamic prediction. By enabling near real-time inference integrated with existing imaging pipelines, it allows direct comparison with hospital phase-diagram assessments and extends them with physically grounded, high-resolution flow fields. This work transforms high-fidelity simulations from an expert-only research tool into a deployable, data-driven decision support system. Our full pipeline delivers high-resolution hemodynamic predictions within minutes of patient imaging, without requiring computational specialists, marking a step-change toward real-time, bedside aneurysm analysis.

Figures (19)

• Figure 1: A) The inputs are defined on an unstructured mesh of, on average, 250,000 nodes. They are comprised of the velocity, the acceleration, the positions and types of the nodes, and finally, statistics about the inflow boundary conditions. In total, this makes for 15 variables per node. The inflow conditions follow the usual values from a cardiac cycle. An example of a mesh with a slice inside the aneurysm bulge is also presented. The model makes a one-step ($\Delta t = 0.01s$) prediction and auto-regressively predicts an entire cardiac cycle ($\approx 0.8s$). B) The pre-training dataset comprises 101 standardized aneurysms from the AnXplore dataset. The fine-tuning dataset comprises 13 real aneurysms, simulated with up to 6 different input waveforms. Finally, the validation set is made of 4 geometries (for a total of 5 aneurysms) from the MATCH challenge. C) The model is trained in 2 main phases (pre-training and fine-tuning), each split into two sub-phases. The first sub-phase sees the model being trained on a coarse version of the meshes. This allows for much faster training steps. The second sub-phase trains the model on the regular (or fine) meshes. During the first sub-phase, we split the training into 2 (both of 150k steps). In the first split, the coarse meshes are masked to make predictions harder and enhance generalization, similar to a Masked AutoEncoder strategy. D) Our model takes the graph nodes' features and the adjacency matrix as inputs. Each transformer comprises masked Multi-Head Self-Attention followed by a Gated MLP with residual connection and Layer Normalization. The Multi-Head Self-Attention masks the node's features by computing a sparse matrix multiplication indexed on the Augmented Adjacency matrix. The adjacency matrix is improved with Dilation, Global Attention, and Random Jumpers. E) The distribution of geometries and waveforms between the fine-tuning dataset and the validation set are highly dissimilar. The validation set exhibits highly out-of-distribution patterns, both in terms of geometry and waveforms. This makes for a challenging task in terms of generalization.
• Figure 2: A) Our model makes a one-step prediction by default. We use this to compute a 1-step error metric. We also use our model in an autoregressive fashion to predict an entire cardiac cycle and compute an All-rollout error based on all time steps from the cardiac cycle. B) We showcase results on the pre-training dataset. We display predicted systolic flow, errors inside the bulge at $y=8$ and $y=10$, a comparison of three selected points at $y=8, y=9, y=10$, and a comparison between the CFD (top row) and our transformer (bottom row) for $v_y$ inside the bulge in a 2D plan defined by $x=0, y=10, z=0$. The GNN model accurately captures both velocity fields and derived wall shear features. Additional visualisation can be found in \ref{['fig:res-pretraining']}. C) Our model outperforms every physical-based GNN on the pre-training dataset on 1-step and All-rollout errors. Results are $\times 10^{-3}$. D) Similar to subfigure B), our GNN follows very closely the ground truth (less than 1% error) of the TAWSS on a full cardiac cycle. The RMSE is in mm/s. E) We showcase results on the validation. We display predicted systolic flow in the artery and the bulge and a plane comparison inside the shapes between the CFD (top row) and our transformer (bottom row). Overall, flow directions are very well predicted, but our model often lacks velocity magnitude (around 10% on average). Additional visualisation can be found in \ref{['fig:res-validation']}.
• Figure 3: A) Velocity contours at systole for the validation cases. The chosen velocities are, from left to right, 600, 100, 200, and 100 mm/s. B) Distribution between CFD and GNN on the validation and test sets from the fine-tuning. We compare them for $\mathbf{v}$, $\mathbf{v}_{max}$, $\mathbf{v}_{min}$, $\text{WSS}$, $\text{WSS}_{max}$, $\text{WSS}_{min}$, $\text{TAWSS}$, $\text{TAWSS}_{max}$, $\text{OSI}$ and $\text{OSI}_{max}$. C) Evolution on the velocity and WSS (following \ref{['eq:delta_metric']}) per time step for all validation cases, plotted as a density map. D) Performance of our model on the fine-tuning dataset and the validation set against two other GNNs approaches. Results are $\times 10^{-3}$. E) TAWSS distribution on the four validation cases and 1 case from the fine-tuning test set. F) Evolution of the velocity and WSS per time step for all validation cases. G) Correlation coefficient between CFD and GNN for several important metrics. We also display the evolution of said metrics as relative changes between CFD and GNN, following \ref{['eq:delta_metric']}
• Figure 4: Cases from the AnxPlore dataset.
• Figure 5: The 13 cases from the Few-shot dataset.