Treatment planning for transcranial ultrasound therapy using deep learning

Lead Research Organisation: University College London
Department Name: Institute of Health Informatics

Abstract

The broad aim of this project is to deliver personalised and targeted treatments for patients undergoing ultrasound therapy in the brain (this is closely connected to one of the three primary aims of the CDT). Deep learning has three potentially transformative roles to play in achieving this:
1. Providing accurate and personalised maps of the skull geometry and material properties. For example, deep learning could be used to map from MR images to CT images, avoiding the need for ionising radiation, to map from CT/MR images to high-resolution images containing the skull microstructure, or to map from CT/MR images directly to maps of the acoustic properties of the skull.
2. Providing real-time calculations of the phase corrections needed to focus the ultrasound waves at the desired target. Currently, these calculations are performed on large supercomputers, and are repeated for each patient and brain target. This means there is a significant delay (at least one week) between acquiring the planning images and selecting the brain target, and delivering the therapy. A deep learning model could be trained by building a lower-order representation of the skull (e.g., using a statistical shape model), and then generating the training set using the computer models mentioned above. This would potentially provide online and adaptive treatment planning, allowing the planning images and therapy delivery to be performed in the same session.
3. Predicting the delivered map of ultrasound dose delivered to the brain. In addition to calculating the phase corrections needed to focus the ultrasound waves through the skull to the desired target, a dose map must also be calculated. This allows the clinical team to review the ultrasound dose delivered to the brain, not only to the target area, but also to the surrounding tissue. A deep learning model could be used to compute these dose maps, e.g., using a residual neural network-based deep learning model.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S021612/1 01/04/2019 30/09/2027
2247858 Studentship EP/S021612/1 01/10/2019 30/09/2023 Joanna ROSALIND Sheppard