Stratifying Intracranial Aneurysms Using Kmeans Clustering on Computational Flow Dynamics Data

Introduction: Intracranial aneurysms carry a high risk of rupture, leading to significant hemorrhage, morbidity, and mortality. Early identification of aneurysms prone to rupture is crucial for timely intervention. Computational Flow Dynamics (CFD) offers insights into aneurysm morphology, but translating these insights into clinical practice is challenging due to morphological diversity. This study integrates unsupervised learning with CFD to enhance rupture prediction by clustering aneurysms based on detailed morphological features.

Methods: Data were sourced from the AneuriskWeb project, utilizing 3D rotational angiography to obtain aneurysm-specific biomechanical features. KMeans clustering was applied to the dataset after dimensionality reduction with t-SNE for clearer cluster separation. Key morphological parameters, such as sac volume, surface area, vessel diameter, and curvature metrics, were used to define clusters. Optimal KMeans hyperparameters were determined via the elbow method, and clustering efficacy was assessed with silhouette scores and variance metrics. Statistical tests, including t-tests and Mann-Whitney U tests, evaluated feature distributions across clusters. Mutual Information (MI) quantified the association between features and cluster labels.

Results: t-SNE successfully reduced dataset dimensionality (n=103), followed by KMeans clustering with k=2, yielding a Silhouette Score of 0.857. MI scores indicated strong associations between cluster labels and features like 'maxRadius', 'vesselDiameter', 'meanRadius', 'tortuosity', and 'maxCurvature'. Cluster 0 ('high-risk') included a greater proportion (53.7%) of ruptured aneurysms, predominantly terminal aneurysms with smaller sac volumes and greater elongation compared to Cluster 1 ('low-risk'). High-risk aneurysms frequently appeared in the ACA and MCA regions, while low-risk aneurysms were mostly located in the ICA and showed greater tortuosity and curvature.

Conclusion : KMeans clustering on t-SNE-reduced aneurysm data effectively stratified aneurysms into high- and low-risk groups. High-risk aneurysms, characterized by specific morphological features and location, underline the potential of combining CFD with unsupervised learning for rupture prediction. Future studies should validate these clusters with live radiographic analyses and integrate clinical parameters for enhanced prediction models.