Cardiac Activation Maps Reconstruction: A Comparative Study Between Data-Driven and Physics-Based Methods

Figures

• 

  Figure 1. Architecture diagrams of (A) the direct activation mapping neural network (DirectMap) and (B) the spatial adaptation of the time-delay neural network (SATDNN-AT).

• 

  Figure 2. (A) Activation time absolute errors and (B) correlation coefficients with respect to dataset size. Each subfigure corresponds to trained neural network with the mentioned number of neurons per layer (nbNeurons).

• 

  Figure 3. Comparison of the computed electrograms at different location of the atria: Exact solution (red line), solution using SATDNN-AT method (EGM_NN, blue line), SATDNN-AT with filter (EGM_NN_Filt, orange line) and FEM-L1 (EGM_FEM, green line). Points correspond to the estimated activation times. Each plot corresponds to a different node of the atria mesh.

• 

  Figure 4. Comparison of (A) absolute errors and (B) correlation coefficients between exact and computed activation times using FEM-L1, DirectMap, SATDNN-AT and SATDNN-AT with filter for the 32 simulations.

• 

  Figure 5. (A) Exact and reconstructed pacing sites using (B) FEM-L1, (C) SATDNN-AT with filter and (D) DirectMap for a test simulation. Numbers are geodesic distances between exact and estimated pacing sites.

• 

  Figure 6. (A) Exact and estimated pacing sites using (B) FEM-L1, (C) SATDNN-AT with filter and (D) DirectMap for 7 different test simulations. Numbers are geodesic distances between exact and estimated pacing sites.

• 

  Figure 7. Evolution of (A) absolute errors and (B) correlation coefficients between exact and estimated activation times with respect to noise level using FEM-L1, DirectMap and SATDNN-AT.

• 

  Figure 8. Exact and estimated electrograms using (A) FEM-L1 (EGM_FEM_Filt) and (B) SATDNN-AT (EGM_NN_Filt) with respect to the added noise level.

• 

  Figure 9. Inflated torso geometry generated by inflating the original geometry by a factor 1.2.

• 

  Figure 10. Bar graphs of the evolution of (A) absolute errors and (B) correlation coefficients with respect to noise added to activation maps. The results correspond to the testing phase.