Insigneo Seminar: Deep learning cardiovascular biomechanics

We are delighted to welcome Dr Choon Hwai Yap, Associate Professor (Reader) at the Department of Bioengineering at Imperial College London, to the Insigneo Institute to give a talk on 'Deep learning cardiovascular biomechanics' on Wednesday 9 September 2026 at 1 pm at our Digital Twin in Healthcare Network  (DT4HT) meeting.

Abstract:

Image-based biomechanics simulations can emulate the mechanical environment and function of cardiovascular system in health and disease. They have provided improved understanding of physiology and pathology, and can be used to predict intervention outcomes for clinical guidance and are the basis of cardiovascular digital twins. Decades of high-quality research have suggested that mechanical forces are influential to cardiovascular biology, and are responsible for maintaining health or causing diseases. However, to date, biomechanical simulations and mechanical biomarkers have not been translated into clinical applications. A reason for this is that simulations are time consuming and are not high throughput enough to support clinical trials that are necessary for clinical adoption. Another reason is that they are skill intensive to perform, which forms an adoption barrier for clinicians.  Deep learning (DL) can resolve these bottlenecks, creating high throughput simulation tools that can be used with little training. I will discuss a few such deep learning biomechanics tools. 

The first is a framework towards real-time prediction of fluid mechanics for cerebral aneurysms. We first used a DL shape generator for augmenting the low sample size of clinical morphology available, utilizing a novel graph Fourier shape encoding. With the shape generator, we generated a large dataset of diverse morphology and performed mass CFD simulations for a large fluid dynamics dataset. Subsequently, we trained a Graph Transformer for wall shear stress prediction, fully supervised by the CFD dataset, which achieved accurate pulsatile WSS predictions (spatial similarity metric, SSIM = 0.981, maximum-normalized relative L2 error =2.84%).

Secondly, I will discuss deep learning cardiac myocardial finite element simulators. Here, I will again describe a graph-based network that is fully supervised by simulations. In this case, we employed a self-supervised cyclic consistency strategy to enhance performance, where a loss function enforced the match between the forward FE step prediction and the prediction of the reverse of that step. The resulting network has similarly high accuracy (99% of nodes had <2mm displacement error, and 90% had <1mm error, pressure prediction correlation with ground truth of R²=0.98).

Brief Biography:

Dr Choon Hwai Yap has a PhD from Georgia Institute of Technology, and trained as a postdoc at the University of Pittsburgh. He was previously an Assistant Professor at the National University of Singapore, and is now an Associate Professor (Reader) at the Department of Bioengineering at Imperial College London. His work is mainly on cardiovascular computational modelling, including image processing for shape and motion extraction from medical image, image-based biomechanics modelling, and more recently, on deep learning image processing and biomechanics modelling.

Website: https://yaplab.github.io/Research.htm

University of Sheffield's Digital Twin for Healthcare Technologies (DT4HT) network