Abstract : In silico models offer effective tools to address cardiovascular diseases and quantitatively analyze clinical data. Recently, many methods have been proposed to blend numerical solvers with machine learning techniques. These approaches hold promise for the patient-specific personalization of models and for the acceleration of their numerical resolution. This minisymposium will offer a forum to discuss the state-of-the-art and future lines of research toward an increasingly effective integration between clinical data and numerical simulations
Organizer(s) : Francesco Regazzoni, Stefano Pagani, Francisco Sahli Costabal, Simone Pezzuto
[04634] Scientific machine learning approaches for many-query problems in cardiovascular applications
Format : Online Talk on Zoom
Author(s) :
Stefano Pagani (MOX, Department of Mathematics, Politecnico di Milano)
Francesco Regazzoni (MOX, Dipartimento di Matematica, Politecnico di Milano)
Luca Dede' (Politecnico di Milano)
Alfio Quarteroni (Politecnico di Milano, EPFL)
Abstract : In many cardiovascular applications, only partial and (possibly) noisy measurements of the phenomenon are available, limiting data-driven techniques in robustly reconstructing quantities of interest. Scientific machine learning approaches compensate for this partial information by integrating physics-based parametric differential models into machine/deep learning models, enabling efficient and accurate solutions to inverse or parameter estimation problems. In this talk, we present some numerical examples of scientific machine learning strategies in computational medicine.
[04639] GPU-Parallel Cardiac Simulation
Format : Talk at Waseda University
Author(s) :
Toby Simpson (Università della Svizzera italiana)
Rolf Krause (Università della Svizzera italiana)
Abstract : The computational cost of current approaches to whole heart simulation make them impractical for research or clinical application.
We compute a complete description of a human heartbeat, including electrophysiology, solid mechanics and fluid dynamics, on a single Graphics Processing Unit (GPU) within a few minutes.
The implementation via a matrix- and mesh-free Finite Volume discretisation, is simple enough to allow patient-specific fitting or provide ground truth data to machine learning algorithms.
[05161] In-silico perivascular flow and transport
Format : Online Talk on Zoom
Author(s) :
Marie Elisabeth Rognes (Simula Research Laboratory)
Abstract : Your brain has its own waterscape: whether you are reading, thinking or sleeping, fluid flows through or around the brain tissue, clearing waste in the process. These biophysical processes are crucial for the well-being and function of the brain. In spite of their importance we understand them but little. In this talk, I will give an overview of mathematical, mechanical and numerical approaches to gain new insight into mechanisms underlying brain clearance.
[04430] Lipschitz Stabilised Autoencoders in Parameter Identification of Dynamical Systems
Format : Talk at Waseda University
Author(s) :
Haibo Liu (Inria Paris)
Damiano Lombardi (Inria Paris)
Muriel Boulakia (Université de Versailles et Saint-Quentin en Yveline)
Abstract : The present work deals with data-driven modelling. Given a set of partial noisy observations of a dynamical system, we investigate using the Lipschitz stabilised auto-encoder to perform an intrinsic dimension estimation to understand how many parameters are responsible for the observed variability. By incorporating the information of the intrinsic dimensionality, we investigate a data-driven model that can complement the classical parameter identification method with the help of data.
01088 (2/3) : 4C @E708 [Chair: Francesco Regazzoni]
[05201] Accelerating hemodynamic predictions via machine learning
Format : Online Talk on Zoom
Author(s) :
Noelia Grande Gutierrez (Carnegie Mellon University)
Abstract : Image-based computational blood flow simulations allow quantifying patient-specific hemodynamics with applications for personalized diagnosis, risk stratification, and treatment selection. However, the clinical translation of these methods is limited due to their high computational cost. We propose machine learning super-resolution to accelerate hemodynamic predictions. For upsampling simulation results, we combine physics-based simulations on a coarse mesh with a graph neural network. Unstructured data (mesh) can be directly transformed into a graph representation, minimizing information loss.
[04443] The fibrotic kernel signature: simulation-free prediction of atrial fibrillation
Format : Online Talk on Zoom
Author(s) :
Francisco Sahli Costabal (Pontificia Universidad Católica de Chile)
Simone Pezzuto (Università di Trento)
Lia Gander (Università della Svizzera Italiana)
Tomás Banduc (Pontificia Universidad Católica de Chile)
Abstract : We propose a fast classifier that is able to predict atrial fibrillation inducibility in patient-specific cardiac models. This is achieved by training the classifier on a variant of the Heat Kernel Signature, which includes information about the fibrosis. These features are fast to compute, when compared to standard cardiac models. The classifier is able to predict the inducibility of single points and also the overall inducibility of the model.
[05176] Learning Reduced-Order Models for Blood Flow Simulations Using Graph Neural Networks
Format : Talk at Waseda University
Author(s) :
Luca Pegolotti (Stanford University)
Martin Pfaller (Stanford University)
Natalia Rubio (Stanford University)
Rita Brugarolas Brufau (Intel )
Ke Ding (Intel)
Eric Darve (Stanford University)
Alison Marsden (Stanford University)
Abstract : We develop one-dimensional reduced-order models for simulating blood flow dynamics in complex cardiovascular geometries using a graph neural network trained on 3D hemodynamic data. Our method, which is a modified version of MeshGraphNet, accurately and efficiently predicts pressure and flow rate with errors below 2% and 3%, respectively, outperforming traditional physics-based models while maintaining high inference efficiency. Our findings demonstrate the potential of this approach in handling diverse anatomies and boundary conditions in physiological settings.
[04920] Fast and accurate reduced order modelling techniques for the simulation of blood flow dynamics
Format : Talk at Waseda University
Author(s) :
Gianluigi Rozza (SISSA Trieste)
Abstract : Heart disease is one of the main cause of death worldwide, therefore in the last years the medical profession
has shown a growing attention for simulating blood flow dynamics through numerical methods.
The main purpose is to build a support for surgical procedure and to predict the progression of a disorder.
Full order mathematical models can be adopted for patient-specific cases, varying physical and geometrical
parameters, however the complexity of the computational domain requires a fine discretization and
as a result a considerable amount of time. Our works focus on the study of Reduced Order Models
(ROMs), which are specifically formulated to reduce the computational cost of complex dynamics such
as biomedical ones. A complete decoupling between an offline and an online stage is adopted to speed up
high fidelity simulations, by splitting what can be done only once and what need to be evaluated for every
new parameter to obtain e good ROM solution. Both intrusive and data-driven approaches are tested for
patient-specific applications to investigate both the efficiency and the accuracy of the ROM framework.
[04420] Parameter estimation in cardiac biomechanical models based on physics-informed neural networks
Format : Online Talk on Zoom
Author(s) :
Federica Caforio (Institute of Mathematics and Scientific Computing, NAWI Graz, University of Graz)
Francesco Regazzoni (MOX, Dipartimento di Matematica, Politecnico di Milan)
Stefano Pagani (MOX, Department of Mathematics, Politecnico di Milano)
Alfio Maria Quarteroni (MOX, Department of Mathematics, Politecnico di Milano)
Gernot Plank (Gottfried Schatz Research Center: Division of Biophysics, Medical University of Graz)
Gundolf Haase (Institute of Mathematics and Scientific Computing, NAWI Graz, University of Graz)
Abstract : In this talk a novel methodology is proposed, based on the integration of physics-informed neural networks methodologies with biophysically detailed three-dimensional cardiac biomechanical models, to generate robust and effective surrogate reduced-order models that are able to reconstruct displacement fields and locally estimate heterogeneous passive mechanical properties. The accuracy and robustness of the proposed method are demonstrated in several benchmarks. This methodology potentially paves the way for the robust and effective identification of patient-specific physical properties.
[03697] Super-resolution and denoising of 4D flow MRI via implicit neural representations
Format : Talk at Waseda University
Author(s) :
Simone Saitta (Politecnico di Milano)
Marcello Carioni (University of Twente)
Subhadip Mukherjee (University of Bath)
Carola-Bibiane Schönlieb (University of Cambridge)
Alberto Redaelli (Politecnico di Milano)
Abstract : We trained sinusoidal representation networks (SIRENs) for denoising and super-resolution of time-varying 3-directional velocity fields measured in the aorta by 4D flow MRI. The performance of different SIREN architectures was evaluated on synthetic measurements and then we applied the best architecture to real 4D flow data of an aortic aneurysm. Our method provides a continuous representation of 4D velocity fields (super-resolution) and achieves denoising thanks to SIREN’s spectral bias, outperforming state-of-the-art techniques.
