PDF e-Pub

Section: New Results

Computational Cardiology & Image-Based Cardiac Interventions

Cardial Electrophysiological Model Learning and Personalisation

Participants : Nicolas Cedilnik [Correspondant] , Ibrahim Ayed [Sorbonne, LIP6, Paris] , Hubert Cochet [IHU Liryc, Bordeaux] , Patrick Gallinari [Sorbonne, LIP6, Paris] , Maxime Sermesant.

This work is funded by the IHU Liryc, Bordeaux.

modelling, electrophysiology, ventricular tachycardia, ischemic cardiomyopathy

This project aims at making electrophysiological model personalisation enter clinical practice in interventional cardiology. During this year:

• we evaluated a fully automated computed tomography-based model personalisation framework in the context of post-ischemic ventricular tachycardia [35],

• we developped a model personalisation methodology based on invasive data in our participation in the STACOM2019 modelling challenge [37],

• we proposed a deep learning based approach to replace numerical integration of partial differential equations used in cardiac modelling [32], see Figure 17.

Figure 17. Transmembrane potential obtained with a reaction diffusion model (top) and forecasted by EP-Net (bottom) for one slice of a tissue slab

Deep Learning Formulation of ECGI for Data-driven Integration of Spatiotemporal Correlations and Imaging Information

Participants : Tania Marina Bacoyannis [Correspondant] , Hubert Cochet [IHU Liryc, Bordeaux] , Maxime Sermesant.

This work is funded within the ERC Project ECSTATIC with the IHU Liryc, in Bordeaux.

Deep Learning, Electrocardiographic Imaging, Inverse problem of ECG, Electrical simulation, Generative Model.

Electrocardiographic imaging (ECGI) aims at reconstructing the electrical activity of the heart using body surface potentials.To achieve this one has to solve the ill-posed inverse problem of the torso propagation. We propose in [33] a novel Deep Learning method based on Conditional Variational Autoencoder able to solve ECGI inverse problem in 2D. This generative probabilistic model learns geometrical and spatio-temporal information and enables to generate the corresponding activation map of the specific heart.

120 activation maps and the corresponding Body Surface Potentials (BSP) were generated using the dipole formulation. 80% of the simulated data was used for training and 20% for testing.We generate 10 propbable solutions for each given input using our model. The Mean Squarre Error (MSE) metric over all the tests was 0.095. As results we were able to observe that the reconstruction performs well. Next, we will extend the model in 3D and test it on real data provided by the IHU Liryc.

Figure 18. Architecture of our conditioned generative model (encoder) and our conditioned variational approximation (decoder)
Figure 19. (a) Simulated and (b) predicted mean activation maps for proposed deep learning based ECGI, (c) Standard deviation map calculated over 10 predictions, (d) error map, difference between predicted and simulated activation maps.

Discovering the link between cardiovascular pathologies and neurodegeneration through biophysical and statistical models of cardiac and brain images

Participants : Jaume Banus Cobo [Correspondant] , Marco Lorenzi, Maxime Sermesant.

Université Côte d'Azur (UCA)

Lumped models - Biophysical simulation - Statistical learning

Figure 20. a) Summary of the available data for each subject, including cardiac data, socio-demographic information, blood pressure measurements and brain volumetric indicators. b) Simplified representation of the lumped model showing the parameters used in the personalisation. $\tau$ characterizes the contractility of the main systemic arteries, ${R}_{p}$ the peripheral resistance, ${P}_{ven}$ the venous pressure right after the capillaries, ${R}_{0}$ the radius of the left ventricle, ${\sigma }_{0}$ the contractility of the cardiac fibers and ${C}_{1}$ their stiffness. A more detailed representation of the myocardial forces is omitted for the sake of clarification. c) Example of the pressure and volume curves that can be obtained from the model, from these curves we extract scalar indicators to match the available clinical data.

Parallel transport of surface deformations from pole ladder to symmetrical extension

Participants : Shuman Jia [Correspondant] , Nicolas Guigui, Nicolas Duchateau, Pamela Moceri, Maxime Sermesant, Xavier Pennec.

The authors acknowledge the partial funding by the Agence Nationale de la Recherche (ANR)/ERA CoSysMedSysAFib and ANR MIGAT projects.

We proposed a general scheme to perform statistical modeling of the temporal deformation of the heart, directly based on meshes. We encoded the motion and the intersubject shape variations, with diffeomorphisms parameterized either by stationary SVFs or by time-varying velocity fields in the LDDMM framework.

Experiments on a 4D right-ventricular endocardial meshes database demonstrated the stability of our transport algorithm, of importance for the assessment of pathological changes. The method is adaptable to other anatomies with temporal or longitudinal data.

Figure 21. Illustration of parallel transport of vectors $a$ and $b$ along a curve (left) and its application to cardiac imaging (right) with a focus on surfaces.

Machine Learning and Pulmonary hypertension

Participants : Yingyu Yang [Correspondant] , Stephane Gillon, Jaume Banus Cobo, Pamela Moceri, Maxime Sermesant.

cardiac modelling, machine learning

Figure 22. The main idea and logic of this work

Style Data Augmentation for Robust Segmentation of Multi-Modality Cardiac MRI

Participants : Buntheng Ly [Correspondent] , Hubert Cochet [IHU Liryc, Bordeaux] , Maxime Sermesant.

Image Segmentation. Multi-modality, Cardiac Magnetic Resonance Imaging, Late Gadolinium Enhanced, Deep Learning

The strategy aims to reduce over-fitting of the network toward any specific intensity or contrast of the training images by introducing diversity in these two aspects, as shown in figure 23.

Figure 23. Different variation of input images and the image processing methods used. C0 denotes the steady-state free precessing CMR modality image.

Towards Hyper-Reduction of Cardiac Models using Poly-Affine Deformation

Participants : Gaëtan Desrues [Correspondant] , Hervé Delingette, Maxime Sermesant.

Model Order Reduction, Finite Elements Method, Affine Transformation, Meshless

Patient-specific 3D models can help in improving therapy selection, treatment optimization and interventional training. However, these simulations generally have an important computational cost. The aim of this project is to optimize a 3D electromechanical model of the heart for faster simulations [38]. The cardiac deformation is approached by a reduced number of degrees of freedom represented by affine transformations (frames in Figure 24b) located at the center of the AHA regions (Figure 24a). The displacement of the material points are computed using region-based shape functions (Figure 24c).

Figure 24. Framework on a cardiac topology. AHA regions (a). Affine degrees of freedom (b). Shape function in one region (c).