Title: Investigating Machine Learning for Rapid Step-by-Step Emulation of Mechanical Thrombectomy Simulations in Ischemic Stroke Treatment

Abstract:

Mechanical thrombectomy for ischemic stroke requires clinicians to make critical decisions under severe time pressure. While numerical physics simulations hold the potential to guide operators toward superior treatment strategies and device choices, their computational latency currently renders them impractical for real-time clinical use. This thesis explores the viability of employing machine learning-based surrogates to replicate these simulations sequentially, aiming to achieve significant acceleration without compromising accuracy.

To evaluate this approach, we trained three distinct surrogate models using data from two simulations featuring simplified aspiration procedures, each characterized by different degrees of geometric complexity. The findings indicate that two of the proposed models can accurately forecast individual simulation steps and deliver considerable performance gains, particularly when augmented with specific data enhancement techniques. Nevertheless, the models demonstrated instability when applied to complex geometries over extended durations. Ultimately, this research establishes a baseline for subsequent investigations aimed at creating stable, scalable methods capable of handling realistic numerical physics simulations of mechanical thrombectomy.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC