Using Deep Learning to Revolutionise Congenital Heart Disease Interventions

Project ID: 2228cd1433 (You will need this ID for your application)

Under Offer

UCL Lead department: Institute of Cardiovascular Science

Project Summary:

The project that we are offering relates to using Deep Learning (DL) to accelerate computational modelling in structural and congenital heart disease. These are diseases in which life changing decisions about therapies (like surgery) must be made purely on the clinician’s experience. Computational modelling can robustly predict a patient’s response to therapy and help decision making, but it is rarely used in clinical practice because the pipeline is too time consuming. We aim to accelerate the whole computational modelling pipeline using DL, including speeding up image segmentation using machine vision methods and rapid modelling (Finite Element Modelling - FEM and Computational Fluid Dynamics – CFD).

We will specifically build on our work in Graph Neural Networks, Convolutional Neural Networks (UNets), unsupervised learning with Autoencoders, and Generative Models to achieve our goals. We have recently developed a pipeline that leverages hybrid image and graph convolutional networks to segment vessels (e.g. the pulmonary arteries) and a novel graph convolutional network to compute blood pressure and velocities. This method is fully automated and speed up the whole process 4000-5000x. Thus, we believe that this work could revolutionise the use of computational modelling in clinical practice.

The student will build on our framework to include time-varying models that can predict pressure and velocity across the cardiac cycle. This will involve investigation of novel new architectures better suited to time varying signals including Graph Transformers and Recurrent Graph Networks. In the PhD, these ideas will be developed to more complex disease and types of computational modelling. This will include simulation of interventions that involve multiple graphs - i.e. predicting if a vessel narrowing (represented by one graph) can be adequately opened by a stent (represented by another graph), which will require significantly more complex DL architectures.