In this work, we solve an inverse biomedical flow problem in 4D with unprocessed, noisy and temporally sparse MRI data on a complex domain. Classical approaches require careful meshing of the brain geometry and making assumptions on the boundary conditions28. In patient-specific brain modeling the meshing is particularly challenging and requires careful evaluation of the generated meshes29. Physics-informed neural networks have been applied for the discovery of unknown physics from data without meshing and without regularization3. This makes the PINN method an appealing and promising approach that avoids major challenges in our application and is therefore well worth investigation. However, PINNs introduce other challenges such as the choice of the network architecture, the optimization algorithm and hyperparameter tuning, e.g., weight factors in the loss function. Nevertheless, it is worth to examine how PINNs perform compared to classical algorithms in our application.
Cathryn Mitchell is Professor of Electronic and Electrical Engineering and EPSRC Research Fellow at the University of Bath. She is interested in all sorts of tomography problems ranging from medical imaging to space physics.
Particle physics has also made possible and advanced other treatment options, including accelerator-based therapy. Each year, tens of millions of patients receive X-ray, proton and ion therapy to treat cancer at more than 10,000 hospitals and medical facilities around the world. 2b1af7f3a8