Cardiovascular Biomechanics and Ultrasound Laboratory
Imperial College London, Department of bioengineering

Menu

Lab Home

People

Research Projects

Publications

Position Vacancy

Machine Learning Biomechanics

Overall Methodology

We employ deep learning networks that are constrained by physics governing equations to extract motion, myocardial strains, and to perform computational flow simulations as well as finite element myocardial tissue mechanics simulations. We aim to have pipelines that start directly from the images for seamless processing.

Accurate CNN Image Registration of Echocardiography Images

We developed a new Convolutional Neural Network (CNN) framework for image registration of echocardiography images, to aid motion tracking and strain computations, that has better performance than existing techniques. Our approach combines shape encoding, adversarial training, and multi-scale registration to achieve this, it could out-perform gold standard traditional techniques such as using the optimal use of algorithms like SimpleElastix.

Our Algorithm, ACDLIR, produced displacement fields that more accurately warped the moving image to match the fixed image than alternative algorithms.

PINN Image Registration of MRI Images

We further developed a new Physics Informed Neural Network (PINN) framework for image registration of echocardiography images, which performs cyclic regularization of fast registration outputs. We preliminarily inserted a Neo-Hookean passive material behaviour model to constraint the network to demonstrate feasibility of physics constraints, and are moving towards more complex physics constraints. The goal is to have a network that perform both motion tracking and finite element modelling of cardiac biomechanics directly from the images.

Our proposed PINN network, Fourier-WarpPINN

Results show that our registration approach enabled better adherence to the incompressible deformation criterion of myocardium, and provided smoother warped forms.

Multi-Case PINN for Vascular Flow Simulations

We are developing a PINN network that is pre-trained in a wide range of possible vascular geometries, by parameterizing the geometries as PINN inputs. This will enable flow simulation results to be very quickly generated when a new case that is unseen by the network is presented. Preliminary work showed that a partial hyperparameter network is the best architecture.

Schematic of the various PINN approach: mixed-input: no incorporation of a hyperparameter network, (dots); fully hypernetwork - a hyperparameter network that determines weights in the PINN network (dashes); modes - a partial hyperparameter network that determined some weights in the PINN network (dot-dashes). Results showed that the modes network was the best at reducing flow velocity and pressure errors in a stenotic vascular simulation.