A Neural Network-Based Surrogate Model for Simulating Transcranial Ultrasound Stimulation

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Transcranial Ultrasound Stimulation (TUS) is a novel brain stimulation technique with unparalleled potential in neuroscience and clinical applications. To reach the unprecedented accuracy that this technique allows, as well as conform to the safety standard, personalized computational simulations are necessary. These are currently performed in accurate but computationally expensive finite-element models, leaving many researchers to not simulate and thus leaving the true potential of TUS unexplored. The current project intends to make simulations accessible and virtually instantaneous so they can become part of everyday research and clinical therapy. As this requires a dramatic speed-up of the time it takes to run these simulations, we propose a surrogate model based on the well established technique of Deep Convolutional Neural Networks. Specifically, various network architectures and loss functions are compared for their performance, in addition to developing a novel technique to assess the performance of the network for the current purpose. The results show that for the current proof of concept, an architecture based on AlexNet gives a decent level of accuracy while resulting in a computational speed up of two orders of magnitude. Therefore, we conclude that the proposed methodology is promising for the time efficient modeling of TUS simulations. This forward model will support the field of TUS research immensely by allowing researchers to use the technique to its full potential while still keeping control over the safety. The current forward model additionally paves the way for an inverse model which will allow for a new approach to TUS planning, which will be greatly beneficial to both research and clinical applications.
Faculteit der Sociale Wetenschappen