Visualisation and Uncertainty in Patient-Specific Whole-Heart Modelling
29 and 30 May 2017
Zuse Institute Berlin, Takustraße 7, 14195 Berlin, Germany
Heart disease is a clinical problem of the first order in Western society. Computational models and simulations that integrate physiological understanding with patient information derived from clinical data have huge potential to contribute to improving our understanding of both the progression and treatment of heart disease. Patient-specific models are currently created from prior anatomical and physiological knowledge that provide a reference framework for describing a patient's heart. This reference model is then augmented with clinical information describing, in part, an individual patient's physiology and pathology. Current models are primarily used for integrating and interpreting clinical data to identify the mechanisms underpinning pathologies, indices of cardiac function and/or treatment response. Despite providing new insights into cardiac physiology and pathology this approach fails to realise the full potential of computational models to not only analyse data but also to make predictive forecasts.
To realise the potential of patient specific modelling, simulation, and prediction and simulation requires the adoption of the mathematics of data assimilation and uncertainty quantification, which has enjoyed great success in meteorology, petroleum engineering and astronomy. This workshop aims to stimulate interdisciplinary collaborations and address this question of how to quantify and visualise uncertainty in medical image analysis and model predictions.
This workshop has been made possible by support from the King's College London — Freie Universität Berlin Funding Programme for Joint Research Workshops.
- Date: 29 and 30 May 2017
- Place: Zuse Institute Berlin, Takustraße 7, 14195 Berlin, Germany (Directions)
- Organisers: Steven Niederer (King's College London) and Tim Sullivan (Free University of Berlin and Zuse Institute Berlin)
Travel and Accommodation
The workshop will take place at the Zuse Institute Berlin (ZIB), Takustraße 7, 14195 Berlin, Germany — the “J”-shaped building highlighted on this map. ZIB can be easily reached by public transport:
- underground line U3 to U-Bahn station Dahlem-Dorf, then a 10-minute walk;
- underground line U9 to S- and U-Bahn station Rathaus Steglitz, then bus X83 to bus stop Arnimallee, then a 5-minute walk;
- of course, other routes are possible — see the website of BVG for routing information.
Berlin is served by two airports, Berlin-Tegel (TXL) and Berlin-Schönefeld (SXF). TXL is about 45 minutes from ZIB by public transport, and is mainly served by full-service airlines; SXF is about 1 hour 10 minutes from ZIB by public transport, and is mainly served by low-cost airlines.
A block of rooms at the reduced rate of 80/110 EUR per night (single/double occupancy) has been reserved at the Seminaris CampusHotel Berlin, less than 5 minutes' walk from the workshop venue. Please contact the organisers for the booking code to use when making your reservation with the hotel.
- Daniel Baum (ZIB)
- Ton Coolen (KCL)
- Jana de Wiljes (U Potsdam)
- Marco Favino (U Svizzera Italiana & U Lausanne)
- Steven Gilmour (KCL)
- Leonid Gouberits (Charité)
- Hans-Christian Hege (ZIB)
- Titus Kühne (DHZB)
- Rashed Karim (KCL)
- Pablo Lamata (KCL)
- Giovanni Montana (KCL)
- Steven Niederer (KCL)
- Gernot Plank (U Med Graz)
- Christof Schütte (FUB & ZIB)
- Julia Schnabel (KCL)
- Alexander Sikorski (FUB & ZIB)
- Tim Sullivan (FUB & ZIB)
- Martin Weiser (ZIB)
- Steven Williams (KCL)
- James Winter (KCL)
- Stefan Zachow (ZIB)
The workshop will commence at 12:50 on Monday 29 May and end at 15:00 on Tuesday 30 May. All talks will take place in the Lecture Hall of the Zuse Institute Berlin, Takustraße 7, 14195 Berlin (the “J”-shaped building highlighted on this map). Dinner on Monday evening will be held at the Alter Krug, Königin-Luise-Straße 52, 14195 Berlin, which is just off the map, immediately west of the intersection of Königin-Luise-Straße and Fabeckstraße.
Click the icon next to a speaker to show or hide the abstract and other supplementary information.
Abstract: Image based modelling is aimed to provide clinicians with precise data replacing invasive measurements, supporting clinicians in medical treatment decision and predicting post treatment outcomes. However, modelling based on clinical imaging data is affected by relatively low time and space resolution resulting in an uncertainty of the input data used in simulations. Uncertainty quantification assessing the propagation of image based uncertainty in simulated parameters is a part of the verification and validation process on the road of translation of modelling approaches into the clinical practice. We represent two examples of the uncertainty analysis during two clinical validation studies for: (1) CFD based analysis of the pressure drops in coarctation (narrowing) of the aorta as a parameter used according clinical guidelines for the treatment decision and (2) CFD-based prediction of hemodynamics after the replacement of the deceased aortic valve. Both studies used phase-contrast magnetic resonance imaging derived anatomy and inlet boundary conditions for simulations. The first study investigated the impact of uncertainty in the reconstructed diameter of the stenosis and measured flow rate on the pressure drop. The second study investigated the impact of the aortic valve prosthesis position (plane and rotation angles) on the transvalvular pressure drop and the aortic wall shear stress.
Abstract: Uncertainty estimation has recently attracted much interest in the medical imaging community, since there is no gold standard available in many image segmentation and registration applications. I will present some of our recent advances in harnessing uncertainty estimates in this domain. For this, I will present a probabilistic framework for nonlinear image registration that allows to estimate uncertainty of the deformation estimates, leading to more accurate classification into AD and healthy control subjects. I will then present our work in discrete optimisation for deformable image registration, which is employed for segmentation propagation of multiple layers of supervoxels (3D superpixels) and is shown to greatly improve on brain segmentation quality.
Biography: Julia Schnabel is Chair in Computational Imaging at the Division of Imaging Sciences & Biomedical Engineering, King's College London. She has graduated with an MSc in Computer Science (Diplom Informatik) from the Technical University of Berlin in 1993, and with a PhD in Computer Science from University College London (UCL) in 1998. She has held a number of postdoc positions at UCL, Utrecht University, NL, and at King's College London, before being appointed Associate Professor in Engineering Science (Medical Imaging) at the University of Oxford, in association with a Tutorial Fellowship in Engineering at St. Hilda's College, Oxford, in 2007 and subsequently becoming Professor of Engineering Science at Oxford in 2014 where she remains a Visiting Professor. Julia's research focuses on complex motion modelling in a range of medical imaging applications, including oncology, cardiology, neurology and perinatal medicine, as well as on developing novel approaches in machine learning.
Abstract: Electromechanical coupling in the heart is a complex multi-physics problem described by a system of PDEs and ODEs, that describes phenomena spanning several temporal and spatial scales.
Due to the complexity of the problem, the construction of a robust discretization schemes and efficient simulation methods is far from being a trivial task.
On the one hand, simulation of electrophysiology with mono- or bi-domain system requires fine spatial meshes and small time-step sizes in order to catch the steep gradients in the action potential and to correctly simulate the stiff ODEs describing the gating variables. While the solution of the spatial problem can be optimally done with algebraic, geometric, or semi-geometric multigrid, time integration still remains the bottleneck for accurate simulations.
On the other hand, solid mechanics can be simulated on coarser meshes but the main difficulty is given by the generalised saddle-point structure of the problem, for which no Lagrangian function exists.
In this talk, we will discuss the derivation and the performances of high order exponential time integration methods for cellular models in electrophysiology and a novel Augmented-Lagrangian, Uzawa-like, multigrid approach for the efficient solution of the mechanical problem.
Biography: Marco Favino holds two half-time post-doctoral positions at the Center for Computational Medicine in Cardiology (Università della Svizzera Italiana) and at the institute of Earth Sciences (University of Lausanne). His main research interests are cardiac modelling and efficient solution methods for it. In particular he is interested in exponential time-integrators and efficient solution methods for saddle-point problems. In Lausanne, he continues his research started during his PhD on poroelasticity, applying this generalized continuum mechanics model to the study of fracture rocks.
Abstract: In this talk we will present some projects that have been carried out over the past few years in the Mathematics for Life and Materials Sciences division at ZIB. The presentation will take place in the Studio da Vinci at ZIB on a curved VR display. All demos will be fully interactive and in 3D. Beside projects that have a clear medical application, projects from biology, biochemistry and flow dynamics will also be shown.
Biography: Gernot Plank is Associate Professor at the Medical University of Graz, with particular research focus on computational biology and supercomputing with applications in biophysical aspects of defibrillation, cardiac arhythmias, cardiovascular haemodynamics, and segmentation and registration techniques for biological applications.
Abstract: Today's understanding of natural processes in form of computational-based models has improved our ability to predict future outcomes of real life dynamics. Yet the associated models often exhibit some form of uncertainty, e.g., with respect to initial conditions or inherent noise. It is possible to reduce such uncertainties via incorporation of data. The broader term “data assimilation” is used when describing this incorporation of observation-based time-ordered information into models that characterize a specific natural phenomena. The development of computationally feasible techniques applicable for the seamless integration of data has been largely driven by the respect fields associated with the dynamic under consideration. Consequently many algorithms have been invented in an ad-hoc fashion and there is a lack of mathematical understanding of the theoretical properties of some of the very popular and widely employed methodologies. However a rigorous mathematical justification of data assimilation approaches is essential to make these techniques universally applicable to phenomena from arbitrary applicational fields. With this goal in mind we recently derived stability and accuracy results for the ensemble Kalman filter, a very commonly employed and robust data assimilation scheme, and future research will include to generalize these findings to a broader class of algorithms.
Biography: During her doctoral research Jana de Wiljes worked on the development of non-stationary non-homogenous data analysis tools at the FU Berlin. She completed her PhD in mathematics under the supervision of Prof. Rupert Klein and Prof. Illia Horenko in December 2014 and currently holds a post-doctoral position with a research focus on data assimilation in Potsdam.
Abstract: Computationally expensive simulations with multiple input settings can be thought of as a type of experiment. The theory and methodology of optimal design of experiments can be adapted and used to maximise the relevant information that can be obtained from a fixed amount of experimental effort. Optimal designs depend on the objectives of the study, the specific models which will be used, how much is known about the system and whether the design can be adapted sequentially, or has to be a one-shot experiment. Some general methods will be described and how specific solutions can be developed for specific applications will be discussed.
Abstract: Solving bidomain or monodomain equations with sufficient accuracy on realistic geometries is still a major computational effort.
This is particularly challenging when several simulations are required as in optimization, inverse problems, and UQ.
In this talk, we explore different options of combining cheap eikonal models with monodomain models in order to achieve better accuracy-effort trade-offs.
In particular we consider eikonal initialization of higher order SDC integrators, heterogeneous domain decomposition, and multilevel optimization.
Joint work with A. Sali and F. Chegini.
Biography: Martin Weiser is head of the Research Group Computational Medicine and the Numerical Mathematics department of the Zuse Institute Berlin. His research interests lie in the area of scientific computing, involving numerical analysis, computer science, applications, and functional analysis. In particular, he is occupied with finite element methods for partial differential equations; applications in medicine; optimal control problems; multi-scale and multi-physics problems; function space oriented numerics; and generic and object oriented implementations.
Biography: Pablo Lamata is a Lecturer and Sir Henry Dale Fellow at King's College of London. His main motivation is to research and develop innovative imaging and modelling solutions for surgeons and physicians, and his goal is to see the results of this effort transferred into the Operating Theatre and Hospital. He has 15 years of experience in the development and clinical adoption of image analysis, physiological modelling, and surgical simulation and navigation solutions. He is currently working at King's College of London, pursuing the clinical adoption of computational models of the heart for a better management of cardiovascular diseases.
Abstract: Starting with an introduction to Bayesian parameter estimation, we will introduce a method for estimating the prior from cohort data. Since the arising non-parametric empirical Bayes estimator leads to unsatisfactory results in the finite data regime, we propose a regularization scheme motivated by information-theoretic constructions resulting in a generalization of reference priors to the empirical Bayes setting. Finally we discuss its application to a high-dimensional systems biology model.
Abstract: Visualisation of high-resolution cardiac data has important applications within heart failure research. The making of two-dimensional maps of three-dimensional objects have a long history going back to the cartographers of the ancient Greeks. In medical imaging, flat maps of the brain have been proposed for Neurology research. The heart presents various challenges. In this talk, I will present some research on a surface flattening approach for visualising the heart as a flat map. A few other visualisation techniques for analysing cardiac data which we have recently used in heart failure research will also be covered.
Biography: Rashed Karim completed a PhD on cardiac image data segmentation under the supervision of Prof. Daniel Rueckert. His current work involves knowledge discovery from high resolution cardiac data using image analysis and visualisation techniques. The applications of his work are in heart failure research. He has also organised three international cardiac data challenges for establishing open-source benchmarks to accelerate algorithm development and testing. Rashed Karim is currently Research fellow at King's College London and Honorary Lecturer at the National Heart and Lung Institute at Imperial College London.