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

Tania Bacoyannis (1, 2) , Julian Krebs (3, 1, 2) , Nicolas Cedilnik (1, 2, 4) , Hubert Cochet (4) , Maxime Sermesant (1, 2)

1 EPIONE - E-Patient : Images, données & mOdèles pour la médeciNe numériquE
2 COMUE UCA - COMUE Université Côte d'Azur (2015-2019)
3 Siemens Healthcare
4 IHU-LIRYC

The challenge of non-invasive Electrocardiographic Imaging (ECGI) is to recreate the electrical activity of the heart using body surface potentials. Specifically, there are numerical difficulties due to the ill-posed nature of the problem. We propose a novel method based on Conditional Variational Autoencoders using Deep generative Neural Networks to overcome this challenge. By conditioning the electrical activity on heart shape and electrical potentials, our model is able to generate activation maps with good accuracy on simulated data (mean square error, MSE = 0.095). This method differs from other formulations because it naturally takes into account spatio-temporal correlations as well as the imaging substrate through convolutions and conditioning. We believe these features can help improving ECGI results.

Mots clés: ECGI; Deep learning; Generative Model; Simulation

Domaines: Modélisation et simulation; Cardiologie et système cardiovasculaire; Apprentissage [cs.LG]

Tania Bacoyannis, Julian Krebs, Nicolas Cedilnik, Hubert Cochet, Maxime Sermesant. Deep Learning Formulation of ECGI for Data-driven Integration of Spatiotemporal Correlations and Imaging Information. FIMH 2019 - 10th International Conference on Functional Imaging and Modeling of the Heart, Jun 2019, Bordeaux, France. pp.20-28. ⟨hal-02108958v2⟩