|
Cardiovascular Biomechanics and A.I. Laboratory |
|
|
Imperial College London, Department of bioengineering |
|
RESEARCH
PROJECTS |
|||||||||||||||||||||||||||||||
|
Links to
Our Codes Use
this link to access codes we developed, we welcome all to
try them and give us feedback on them. |
||||||||||||||||||||||||||||||||
|
Deep
Learning Cardiac Imaging and Geometric Modelling We
are developing novel deep learning techniques to improve segmentation and
motion tracking of the heart from echo and MRI images. This includes
strategies to accurately scale up to 3D images, and using motion-enhancement
to improve segmentation.
To
enhance cardiac geometric modelling, we developed next generation algorithms
for 3D reconstruction of cardiac geometries images that can be directly
supervised by image labels, which has superior performance than PCA
statistical shape models and marching cubes. The algorithm can be used to
enhance future deep learning mesh reconstructions from images.
We
recently developed a new approach for reconstructing 3D mesh of objects from
voxelization, the Faithful Contouring Method, that can replace the
age-old Marching Cubes approach. Faithful Contouring uses unsigned
distance field rather than signed distance field for the reconstruction, and
can thus reconstruct objects with multiple zones and surfaces, it does not
require water-tightening and can thus reconstruct single surface objects, and
it achieves near lossless fidelity, effectively preserving sharpness and
internal structures that are challenging for marching cubes. The technique
can be very useful in 3D learning tasks such as mesh generative AI.
Deep
Learning Cardiovascular Fluid Mechanics To
bring biomechanics research to the next level, we are currently developing
deep learning algorithms to generate results of cardiovascular flow
simulations and cardiac finite element simulations, to accelerate the
adoption of biomechanical analysis in clinical practices for disease
detection, evaluation and prognosis. For decades, much scientific evidence
has gathered on how fluid stress parameters such as wall shear stress, OSI,
RRT, etc., affects vascular biology and can dictate disease pathogenesis and
progression, but to date no clinician use them in routine work because
current CFD simulations are too slow to properly support clinical trials. Our
Deep Learning fluid mechanics predictors can change that, giving high
throughput fluid stresses prediction, and allowing clinicians to access fluid
mechanics parameters without specialized skills. In
vascular fluid mechanics, we developed a network for 1-shot network for rapid
prediction of 3D blood flow wall shear stresses for pulsatile flow, and
demonstrated its accuracy for diverse bifurcating cranial aneurysm cases. We
started with a mere 100+ geometries extracted from clinical images. We first
developed a realistic shape generator (MICCAI 2025), which we used to
generate 14,000 realistic shapes. We performed steady CFD flow simulations
for 14,000 cases, and pulsatile CFD flow simulations for 800 cases (dataset
in NeurIPS 2025), which becomes sufficient for
training our WSS prediction network. We then used a GPS graph transformer
network for flow prediction, which achieved high accuracy. Since pulsatile
CFD simulations are slow and expensive, while steady flow CFD is extremely
cheap, we adopted the strategy of using steady flow results to augment the
training for pulsatile flow prediction, which improves results.
We
also tested using Physics Informed Neural Networks (PINN) for cardiovascular
fluid mechanics predictions. However, our conclusion is that PINN is not
ready for real-time prediction with diverse geometric inputs, as convergence
is slow and computational costs are high. Nonetheless, PINN can still be
useful for inverse computation tasks, such as predicting the flow field
from sparse Doppler signals
Deep
Learning Cardiac Myocardial Biomechanics Tapping
on our experience with finite element (FE) modelling of cardiac biomechanics,
we developed deep learning surrogate for computing cardiac biomechanics.
Patient-specific FE simulations of the heart based on clinical images can be
used for predicting cardiac function during disease, when disease severity
progress, or after various treatment options, and this can be extremely
useful for informing treatment decisions. They are the basis of the Cardiac
Digital Twin technology for personalized optimized care. However, FE
simulations are very slow, and cannot properly support a clinical trial to
enable clinical translation. Our Deep Learning FE surrogate simulators can
change that, enabling high throughput cardiac motion and function prediction
without requiring specialized skills from users. We
developed a image-based
PINN surrogate for computing myocardial biomechanics, which can also be used
for back-computing the myocardial stiffness and contractile force. Using
shape modes derived from the individual heart's motions, our reduced order
PINN only need <10min to compute for each new patient.
However,
10 min is still a long time. We have thus developed a graph-attention network
for rapid 1-shot prediction, trained by massive FE simulations. We adapted a
cyclic consistency strategy that greatly enhanced accuracy while allowing a
reduction in training samples. Using this network, we first show that it can
be used to compute the load-free state of the heart from the imaged
end-diastolic state. We have now moved on to using this network for
predicting the entire cardiac cycle FE results (will be published soon)
Improving
Echocardiography Heart Function Evaluation Echocardiography
evaluation of heart function is widely performed, and is important for
diagnosis and to determine if interventions should be performed. However,
many current heart function parameters have
shortcomings, and we have proposed ways improvements: -
the
Ejection Fraction (EF) parameter is widely used to evaluate cardiac health,
but it is a poor indicator when the heart undergoes geometric remodelling
during disease. We proposed a correction for EF to resolve this, and showed
that our new corrected parameter have stronger prognosis capabilities for
rehospitalization. -
in
fetal echocardiography, strain measurements in the literature have widely varying
results, and lack precision. We performed careful measurements of 2D versus
3D cardiac strains from echocardiography, and show the essential reasons for
their disagreement, and point out factors that potentially cause the
variability. -
Microscopy
demonstrated that the myocardium has microscopic sheetlet structures, and it
is hypothesized that sheetlet sliding is important to cardiac function. We
showed via simulations that sheetlet sliding is some effects on function in
the normal heart, but the effects are amplified in hypertrophic hearts. Use
these links below for details.
Biomechanics
of the Human Fetal Heart and Fetal Heart Intervention Biomechanical
stimuli are important stimuli for proper fetal heart development, but we
understand very little of it. The biomechanical environment of the fetal
heart, its growth and remodelling in response to abnormalities, and the
mechanobiological mechanism responsible for malformations are all not
well-understood. Abnormalities during mid-gestation that disrupts the normal
biomechanical environment can lead to congenital heart malformations. In some
such cases, catheter-based intervention on the fetal heart can correct the
abnormalities to prevent the malformation at birth. There is much room for
such interventions to improve, and biomechanics modelling can help such an
effort. We use a range of techniques to study the fetal heart and fetal heart
intervention, including fetal echocardiography image processing,
computational fluid dynamics simulations, and finite element modelling of
myocardial mechanics. Use
these links below for details.
Embryonic
Heart Biomechanics The
embryonic heart is the first organ to develop. It undergoes a fascinating
developmental process, starting out as a simple tube and develops into a
4-chamber structure by week 8 of gestation, and sustaining a tremendous
amount of growth and highly dynamic remodelling. We hypothesize that
mechanical forces are important stimuli to proper early cardiac development,
seek to understand the biomechanics of embryonic hearts of both normal small
animal embryos and those of animal embryonic models of congenital malformations,
as well as understand the mechanobiological pathways towards these
malformations. We use advanced techniques of image-processing and image-based
simulations to obtain greater details in our studies. Use
these links below for details.
Materials Technology Towards a
Blood Pump with Low Blood Damage Blood
pumps save countless lives every day, and include the implanted type (LVAD),
those in the ICU (ECMO), and those in the surgical suite (heart-lung
machine). However, they impose high stresses on blood and induce foreign
surface reactions to cause thrombosis, and thromboembolic complications. We
pursue various strategies and technologies to attain a blood pump with low
blood damage. For example, we fabricate superhydrophobic and superhemophobic surfaces to enable slip flow in blood
pump surfaces to reduce stresses, we seek new ways of pumping blood, such as
using electro-active polymers and utilizing resonance in roller pumping. Use
this link below for details.
Superhydrophobic Hemostatic
Materials Technology Traditional
hemostatic devices rely on absorbing blood to bring
about clotting and hemostasis. We recently
discovered that a nanofibrous superhydrophobic material provides an
alternative and excellent approach towards hemostasis.
The material is strongly repellent towards blood, and easily prevents blood
loss, which is the #1 reason for death in serious injuries. However, the
nanofibers can still cause fast clotting to bring about fast hemostasis. Since the material did not wet while blood
clots, blood is only connected to the path via nano-contacts after clotting,
and this enables extremely easy detachment of the patch from the wound.
Finally, the material is also repellent to microbial attachment and thus has
natural antimicrobial properties. We are currently enhancing this technology,
and pursuing commercialization. Use
this link below for details.
Placenta
and Placenta Disease Biomechanics The
placenta is an important organ during pregnancy, whose health has great
short- and long-term impact on the health of both the mother and the child.
Pregnancy diseases such as Intrauterine Growth Restriction have surprisingly
high prevalence and consequent mortality and morbidity, even in developed
countries, and there is no proven method to prevent or treat the disease. We
advocate that biomechanical approach to studying the placenta can provide new
insights that can lead to better detection, diagnosis, and even treatment.
Examples of our approach include mechanical testing and constitutive
modelling on normal and diseased human placenta samples, investigating the
use of elastography to detect placenta diseases, and image-based biomechanics
simulations on placenta and umbilical blood vessels in health and disease. Use
this link below for details.
|
||||||||||||||||||||||||||||||||
|
|
Collaborations We are very collaborative in our work. If you have
questions about our work, or if we can help you in your research work, please
feel free to contact us. |