From the cluster to the clinic: Real-time treatment planning for transcranial ultrasound therapy using deep learning (Ext.)

Lead Research Organisation: University College London
Department Name: Medical Physics and Biomedical Eng

Abstract

This is an extension of the Early Career Fellowship: Model-Based Treatment Planning for Focused Ultrasound Surgery.

Brain disorders present a huge challenge for health services across the world, with studies showing these conditions affect as many as one third of the adult population. In the UK, approximately 1 in 6 people are affected by a neurological disorder and 1 in 6 by a common psychiatric disorder. The total annual cost of these conditions is estimated to exceed £100 billion. These disorders can be devastating for patients and greatly reduce their quality of life. Today, patients are often treated by the prescription of drugs that alter the way the brain functions. For many patients, this causes a reduction in their symptoms. However, if these drugs are used for long periods of time, their effectiveness often decreases and there can be many side-effects. It can also be difficult for drugs to exit the blood-stream and enter the brain as desired because of a protective lining called the blood-brain barrier. Depending on their diagnosis, some patients may also be offered surgical procedures to remove part of the brain or implant small wires that use electricity to stimulate brain cells.

One exciting alternative to drugs and surgery is the use of ultrasound. Ultrasound imaging is well known for taking pictures of developing babies during pregnancy. However, ultrasound is now also starting to be used to treat brain disorders. This is possible because ultrasound waves cause mechanical vibrations that affect the brain in different ways. For example, they can cause the tissue to heat up or generate forces that agitate the brain cells and tissue scaffolding. Several different types of treatment are possible depending on the pattern of ultrasound pulses used. This includes precisely destroying small regions of tissue, generating or suppressing electrical signals in the brain, or temporarily opening the blood-brain barrier to allow drugs to be delivered more effectively. These treatments are all completely non-invasive and have the potential to significantly improve outcomes for patients.

A major challenge for ultrasound therapy is ensuring the ultrasound energy is delivered to the precise location identified by the doctor. This is difficult because the skull bone is very rigid and causes the ultrasound waves to be reflected and distorted. It is possible to predict and correct for these distortions using powerful computer models of how ultrasound waves travel through the body. However, these models can take many hours or days to run on large supercomputers, so cannot currently be used for patient treatments.

The aim of this fellowship extension is to develop a new type of model that can make very fast predictions of how sound waves travel in the brain. This will be based on a special type of artificial intelligence called deep learning. The deep learning models will be trained to predict the distortion caused by the skull bone. The models will learn this using a large number of training examples generated using the powerful computer models mentioned above. As part of the project, the models will be rigorously tested using patient data from previous clinical treatments. Carefully planned laboratory experiments using phantom materials designed to mimic the skull and brain will also be conducted. The new models will allow doctors to automatically correct for distortions caused by the skull and quickly predict the treatment outcomes. This would be a major breakthrough in the treatment of brain disorders and enable the wide-spread application of ground-breaking ultrasound therapies.

Planned Impact

The direct beneficiaries of this project are patients suffering neurological and psychiatric brain disorders. This covers a wide spectrum of conditions, including Parkinson's disease, Alzheimer's disease, essential tremor, and depression. These disorders are extremely debilitating and have a significant impact on quality of life for patients and carers. Taken together, these conditions comprise the largest single cause of morbidity in the EU in terms of disability adjusted life years. This has clear implications for healthcare budgets and the economy more broadly.

In the last decade, new treatments for brain disorders based on therapeutic uses of ultrasound have generated huge excitement in the research and medical communities. Ultrasound offers the unique ability to non-invasively ablate brain tissue, deliver drugs, stimulate or modulate brain activity, and open the blood-brain barrier. However, one major barrier to the wider clinical adoption of this technology is the lack of accurate online treatment planning tools. The skull can significantly distort the ultrasound waves as they propagate into the brain, so planning tools are essential to correct for these distortions and predict treatment outcomes ahead of time. However, even with large supercomputers, existing treatment planning models can take hours or days to run, making them unsuitable for online use in many clinical applications.

The novel tools for treatment planning based on deep learning outlined in this proposal could provide a major breakthrough in computing performance and act as a catalyst for the widespread clinical application of therapeutic ultrasound technologies in the brain. Impact will arise from: (i) the enhanced accuracy compared to existing models used in commercial devices, (ii) the unprecedented levels of computational performance which will allow model-based treatment planning predictions to be made in real-time, (iii) the extensive validation of the models, and (iv) adherence to the regulatory framework required for the clinical application of scientific software. In the context of delivering value-based healthcare, these tools could also play a significant role in decreasing procedural costs and optimising clinical outcomes. This impact will be enhanced by open-source software releases and the establishment of a new subject repository for ultrasound metrology data.

The enhanced capabilities for ultrasound therapy offered by real-time treatment planning software will provide a significant competitive edge over other planning approaches currently used in academia and industry. This will make the software commercially attractive to manufacturers of therapeutic ultrasound equipment, two of whom are already directly engaged with this project. It is expected the generated IP will lead to licensing agreements or the development of new start-ups, with the UK becoming a base for future international investment. The developed tools will also act as a platform technology for wide-reaching investigations into the interaction of ultrasound with the human body.

Publications

10 25 50

publication icon
Stanziola A (2023) Transcranial ultrasound simulation with uncertainty estimation in JASA Express Letters

publication icon
Stanziola A (2023) A learned Born series for highly-scattering media in JASA Express Letters

publication icon
Roberts M (2023) open-UST: An Open-Source Ultrasound Tomography Transducer Array System. in IEEE transactions on ultrasonics, ferroelectrics, and frequency control

publication icon
Miscouridou M (2022) Classical and Learned MR to Pseudo-CT Mappings for Accurate Transcranial Ultrasound Simulation. in IEEE transactions on ultrasonics, ferroelectrics, and frequency control

publication icon
Kleparnik P (2022) On-the-Fly Calculation of Time-Averaged Acoustic Intensity in Time-Domain Ultrasound Simulations Using a k-Space Pseudospectral Method. in IEEE transactions on ultrasonics, ferroelectrics, and frequency control

publication icon
Jaros M (2020) k-Dispatch

publication icon
Brown M (2020) Stackable acoustic holograms in Applied Physics Letters

publication icon
Brown M (2023) Binary Volume Acoustic Holograms in Physical Review Applied

publication icon
Brown M (2022) Binary volume acoustic holograms in The Journal of the Acoustical Society of America

publication icon
Aubry JF (2022) Benchmark problems for transcranial ultrasound simulation: Intercomparison of compressional wave models. in The Journal of the Acoustical Society of America

 
Description The aim of this proposal was to develop new types of computer modelling tools based on deep-learning that allow us to predict the path of ultrasound waves through the human brain. We have developed a new type of fully-learned solver that allows the distortion due to the skull to be rapidly predicted (https://doi.org/10.1016/j.jcp.2021.110430). We have developed a open-source coding framework for writing differentiable numerical simulators with arbitrary discretizations (https://arxiv.org/abs/2111.05218), which was used as the basis for a new open-source wave simulation toolbox, j-Wave (https://doi.org/10.1016/j.softx.2023.101338). j-Wave is a library of simulators for acoustic applications that can be used as a collection of modular blocks that can be easily included into any machine learning pipeline. j-Wave was validated as part of a major international benchmarking effort led by UCL (https://doi.org/10.1121/10.0013426), and used to study the uncertainty in simulations through the skull (https://doi.org/10.48550/arXiv.2212.04405). We developed a new approach for rapidly solving the Helmholtz equation based on a learned Born series (https://arxiv.org/abs/2212.04948), and learned methods for mapping from MRI image to pseudo-CT images for treatment planning (https://doi.org/10.1109/TUFFC.2022.3198522).
Exploitation Route The models we are developing are likely to be of interest to other research groups and industry.
Sectors Digital/Communication/Information Technologies (including Software),Education,Healthcare,Manufacturing, including Industrial Biotechology

 
Description Capital Award for Core Equipment at UCL
Amount £650,000 (GBP)
Funding ID EP/T023651/1 
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Public
Country United Kingdom
Start 11/2019 
End 05/2021
 
Description Spectral element methods for fractional differential equations, with applications in applied analysis and medical imaging
Amount £103,887 (GBP)
Funding ID EP/T022280/1 
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Public
Country United Kingdom
Start 07/2021 
End 06/2024
 
Description UCL EPSRC IAA 2022-25 FUNDING
Amount £87,417 (GBP)
Funding ID EPSRC IAA 2022-25 KEI2022-02-03 
Organisation Engineering and Physical Sciences Research Council (EPSRC) 
Sector Public
Country United Kingdom
Start 02/2023 
End 01/2024
 
Description UCL Knowledge Exchange and Innovation Fund
Amount £86,087 (GBP)
Funding ID EPSRC IAA 2017-20 Discovery-To-Use 
Organisation University College London 
Sector Academic/University
Country United Kingdom
Start 01/2021 
End 12/2021
 
Title Ramped V1 transcranial ultrasonic stimulation modulates but does not evoke visual evoked potentials 
Description Raw EEG data for the study "Ramped V1 transcranial ultrasonic stimulation modulates but does not evoke visual evoked potentials" 
Type Of Material Database/Collection of data 
Year Produced 2022 
Provided To Others? Yes  
Impact Raw data 
URL https://osf.io/rbcfy/
 
Title TPO Control Toolbox 
Description Third-party toolbox for controlling the Sonic Concepts Transducer Power Output (TPO) systems using a USB connection and MATLAB. The toolbox functions are based on MATLAB codes provided by Sonic Concepts, Inc (distributed under an MIT license), extended to give a uniform interface, add documentation and error checking, and to specify all input parameters in base SI units (Hz, s, etc). The toolbox contains a series of functions for connecting to the TPO (using serial commands over USB), setting the TPO parameters, and triggering the TPO output. 
Type Of Technology Software 
Year Produced 2022 
Open Source License? Yes  
Impact This toolbox was used for two human transcranial ultrasound stimulation studies. 
URL https://github.com/ucl-bug/tpo-control-toolbox
 
Title j-Wave: Differentiable acoustic simulations in JAX 
Description j-Wave is a library of simulators for acoustic applications. Is heavily inspired by k-Wave (a big portion of j-Wave is a port of k-Wave in JAX), and its intended to be used as a collection of modular blocks that can be easily included into any machine learning pipeline. Following the philosophy of JAX, j-Wave is developed with the following principles in mind: to be differentiable, to be fast via jit compilation, easy to run on GPUs, easy to customize. 
Type Of Technology Software 
Year Produced 2022 
Open Source License? Yes  
Impact j-Wave was used in a recent modelling intercomparison, and to study uncertainty in transcranial ultrasound simulation. 
URL https://github.com/ucl-bug/jwave
 
Title jaxdf - JAX-based Discretization Framework 
Description jaxdf is a JAX-based package defining a coding framework for writing differentiable numerical simulators with arbitrary discretizations. The intended use is to build numerical models of physical systems, such as wave propagation, or the numerical solution of partial differential equations, that are easy to customize to the user's research needs. Such models are pure functions that can be included into arbitrary differentiable programs written in JAX: for example, they can be used as layers of neural networks, or to build a physics loss function. 
Type Of Technology Software 
Year Produced 2022 
Open Source License? Yes  
Impact jaxdf is being used by researchers interested in differentiable models, and forms the basis for the j-Wave acoustic simulation software. 
URL https://github.com/ucl-bug/jaxdf
 
Title k-Plan: Ultrasound Therapy Planning 
Description k-Plan is an advanced modelling tool for precision planning of transcranial ultrasound procedures. It uses a streamlined and intuitive workflow that allows users to select an ultrasound device, position the device using a template or medical image, and specify the sonication parameters. High-resolution calculations of the ultrasound field and temperature inside the skull and brain are then automatically calculated in the cloud with a single click. 
Type Of Technology Software 
Year Produced 2022 
Impact k-Plan is developed by researchers at University College London and the Brno University of Technology based on more than a decade of cutting-edge research into ultrasound modelling and planning for transcranial ultrasound therapy. It is the first software tool for model-based treatment planning for ultrasound therapy, and is being brought to market in collaboration with Brainbox, Ltd. 
URL https://k-plan.io/