Context
This paper was my Master Thesis for the Master AI at the University of Amsterdam and is hosted on Arxiv here. I collaborated with the Amsterdam University Medical Centre where my supervisor Thijs Kuipers was based.
Context
This paper was my Master Thesis for the Master AI at the University of Amsterdam and is hosted on Arxiv here. I collaborated with the Amsterdam University Medical Centre where my supervisor Thijs Kuipers was based.
Summary
This paper is a first step towards solving the problem of tool selection when a blood clot is mechanically removed from the brain. As of 2026, when a medical operator decides what tool to use in the operation they choose based on personal preference and on the tool availability, leaving room for improvement. Luckily, numerical computer simulations of these procedures exist, which could aid in tool selection by showcasing likely outcomes when using various tools. However, there is a catch, these simulations are very slow, taking up to a day to run on cluster of computers. This is too long, as it is critical to remove blood clots as fast as possible as the faster the operation takes place the higher the survival and recovery rate. In practice, there is only a 20-60 minute time window between imaging of the brain and the operation. In this paper we attempt to tackle this problem by applying Machine Learning to provide a surrogate to these simulations that is fast enough to be practical.
Summary
This paper is a first step towards solving the problem of tool selection when a blood clot is mechanically removed from the brain. As of 2026, when a medical operator decides what tool to use in the operation they choose based on personal preference and on the tool availability, leaving room for improvement. Luckily, numerical computer simulations of these procedures exist, which could aid in tool selection by showcasing likely outcomes when using various tools. However, there is a catch, these simulations are very slow, taking up to a day to run on cluster of computers. This is too long, as it is critical to remove blood clots as fast as possible as the faster the operation takes place the higher the survival and recovery rate. In practice, there is only a 20-60 minute time window between imaging of the brain and the operation. In this paper we attempt to tackle this problem by applying Machine Learning to provide a surrogate to these simulations that is fast enough to be practical.
In this paper we train three different types of ML models on two simulation datasets constructed at the Amsterdam UMC. These datasets which take the form of point clouds, represent blood clots stuck in either a simplified or a complex blood vessel system. Moreover, we investigate various data augmentations: rotations, point cloud density and generating additional data.
In the end, we find that two of the tested models perform well in simplified vessel geometry even on longer timescales. However, we also find that none of the methods presented are able to perform well on longer timescales for the more complex case, even though they do work well on shorter timescales. We also find that certain data augmentation are crucial for good performance for specific models.