Full Waveform Inversion for Medical Ultrasound Tomography in Julia on multi-xPUs
Open access
Author
Date
2023-10-30Type
- Master Thesis
ETH Bibliography
yes
Altmetrics
Abstract
This Master's Thesis aims to develop efficient yet user-friendly tools to perform seismic tomography using the Full Waveform Inversion (FWI) approach, as derived from the field of Earth sciences, and apply them in the realm of ultrasound medical tomography.
Since its introduction in geosciences a few years ago, FWI has provided spectacular images of the Earth's subsurface at different scales, revealing, with unprecedented details, the internal structure of our planet.
Recently, FWI has also started to be used to reconstruct high-resolution medical images of soft tissues from ultrasound data. However, due to its high computational cost and complexity, it has yet to see extensive use in real-world applications.
This work aims to fill the gap between theory and practice by providing efficient, easy-to-use, and scalable finite-difference-based solvers for acoustic FWI written in the high-level Julia programming language. This allows also non-expert users to perform numerical experiments to test different setups and algorithms and to address real applications with ultrasound data.
Our solvers are device-agnostic and simulations can be distributed on multiple devices (multi-xPUs), providing the user with a range of parallelization options that can fit different problems.
The correctness of the solvers is checked using rigorous tests and synthetic inversions, where we show the potential and pitfalls of the method.
Finally, as an application, a case study inversion with ultrasound data gathered in a medical imaging setting is performed, to assess the feasibility in real-world scenarios.
A set of benchmarks testing the memory throughput, since the algorithms under study are memory-bound, show that our solvers achieve a very high percentage of peak memory throughput (up to 90%) on modern GPUs and good weak scaling on distributed systems.
We conclude that using modern advances in software and hardware provided by the scientific computing community, the computational challenges of FWI can be addressed
to make it a feasible method for ultrasound medical tomography. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000668605Publication status
publishedPublisher
ETH ZurichSubject
Ultrasound Computed Tomography; Full-Waveform Inversion; GPU computingOrganisational unit
03971 - Fichtner, Andreas / Fichtner, Andreas
More
Show all metadata
ETH Bibliography
yes
Altmetrics