Fully Automated Electrophysiological Model Personalisation Framework from CT Imaging

Abstract

There has been a recent growing interest for cardiac computed tomography (CT) imaging in the electrophysiological community. This imaging modality indeed allows to locate and assess post-infarct scar heterogeneity, allowing to predict zones of abnormal electrical activity and even personalise EP models. To this end, most of the literature uses manually segmented CT images where one fundamental information is extracted, the myocardial wall thickness. In this paper, we evaluate the impact of using an automated deep learning (DL) methodology to segment the left ventricular wall and extract relevant scar information on the resulting personalised models. Using CT images from 8 patients that were not used during the DL training, we show that the automated segmentation is very similar to the manual one (median Dice score: 0.9). Thickness information obtained this way is also very close to the manual one (median difference: 0.7 mm). A wavefront propagation model personalisation framework based on this thickness information does not show relevant differences in its output (median difference in local activation time: 2 ms), proving its robustness. Bipolar electrograms, simulated through a novel approach, do not differ significantly between manual and automated segmentations (Pearson's r: 0.99).

Deep learning formulation of electrocardiographic imaging integrating image and signal information with data-driven regularization

“…with σ E the local electrical conductivity, ∇V the spatial gradient of the potential V and ∇T the gradient of the activation map given by the Eikonal model [14]. We can therefore directly compute the local dipole from the gradient of the activation map and the time derivative of a transmembrane potential model (Mitchell-Schaeffer).…”

“…The developed method was used to predict activation maps for different cardiac geometries with their corresponding simulated BSP. The model was trained and tested on a simulated database using a personalised cardiac simulation pipeline [14]. Compared with the classical mathematical formulation of ECGI, this approach has five main advantages:…”

Scar-Related Ventricular Arrhythmia Prediction from Imaging Using Explainable Deep Learning

“…To start with, we generated the ventricular masks from cardiac CT using a pre-trained Dual-UNet segmentation network proposed in [4]. The required ventricle masks included the right ventricular epicardium (RVEPI), the left ventricular epicardium (LVEPI) and endocardium (LVENDO).…”

“…The python package implemented for [3] was made available by the author 3 . A 3D surface model of the LV was computed using the discrete marching cubes algorithm provided by the VTK package 4 and the thickness information was projected onto the mesh.…”

Personal-by-Design: A 3D Electromechanical Model of the Heart Tailored for Personalisation

“…1e). In order to simulate this electrical activity, we consider each voxel of the image as a dipole of current density j eq = −σ∇v where ∇v is the spacial gradient of the potential v [4]. By using the chain rule, we obtain:…”