A Neural Network Approach for Reconstruction of Cardiac Electroanatomical Maps

Abstract

In many engineering and medical applications, local sensors can only provide a partial view of a phenomenon’s geometry and its physical behaviour. Ensuring a high-fidelity reconstruction of both is key for developing accurate data-model integration and statistical learning methods, especially in scenarios where no additional data acquisition modalities are available. In this talk, we present a novel approach that employs neural networks (NNs) to reconstruct a 3D object’s geometry and distributed spatio-temporal quantities of interest, considering the clinical application of electroanatomical (EA) maps for treatment of cardiac atrial fibrillation. A EA map consists of noisy and sparse 3D points and associated electrograms measured by a catheter and is employed to guide electrophysiologists in navigating the geometry and clinical properties of heart chambers. A first model is used for geometry reconstruction by learning the implicit representation of 3D geometries given by their associated Signed Distance Functions. For its training process, we use a combination of loss terms employing geometric pointwise data in an unsupervised or semi-supervised manner and including suitable data fidelity and geometric regularizations terms. A second model is then optimized to predict spatio-temporal fields over the reconstructed geometry. We guarantee that the reconstructed phenomenon is physically plausible by introducing a set of spatial constraints that exploits directly the differentiability of the geometry implicit representation, embedded in the first NN. We provide numerical results to verify the proposed deep learning approach on a benchmark spherical geometry and on realistic electroanatomical data. We achieve a relative error of less than 1% on the reconstructed sphere radius and a faithful differentiable geometrical reconstruction of the patient, that is validated against an alternative imaging technique. The scalar field and its gradient approximation present a reconstruction error of 2% on synthetical data and of 5% on the EA data. The electrogram signal reconstruction model predicts the overall trend on the clinical data, and demonstrates high accuracy on the benchmark dataset. Numerical results show that our method generates accurate and differentiable 3D reconstruction of the left atrium and its electrophysiological properties.