Research on noninvasive electrophysiologic imaging based on cardiac electrophysiology simulation and deep learning methods for the inverse problem

Table 3 Results of TMP reconstruction of cardiac surface for 8 regions by 4 methods

		PSO-BP ^a	1D-CNN ^b	2D-CNN ^c	LSTM ^d
LA	R^{2 e} (%)	93.2 ± 1.40	94.4 ± 0.64	94.3 ± 0.66	96.9 ± 0.40
LA	MAE ^f (mV)	1.00 ± 0.14	0.75 ± 0.05	0.73 ± 0.05	0.40 ± 0.04
RA	R² (%)	85.0 ± 3.80	96.0 ± 0.30	95.2 ± 0.40	96.6 ± 0.35
RA	MAE (mV)	1.57 ± 0.23	0.70 ± 0.04	0.72 ± 0.05	0.45 ± 0.06
LV	R² (%)	84.6 ± 3.80	98.5 ± 0.44	98.6 ± 0.58	99.7 ± 0.11
LV	MAE (mV)	1.88 ± 0.17	0.86 ± 0.40	0.54 ± 0.08	0.22 ± 0.05
RV	R² (%)	90.0 ± 0.90	98.3 ± 0.41	98.7 ± 0.26	99.8 ± 0.03
RV	MAE (mV)	1.57 ± 0.11	0.66 ± 0.04	0.51 ± 0.04	0.19 ± 0.03

^a PSO-BP = Particle swarm optimization-back propagation neural network, ^b 1D-CNN = One-dimensional convolutional neural network, ^c 2D-CNN = Two-dimensional convolutional neural network, ^d LSTM = Long short-term memory network, ^e R² = The coefficient of determination, ^f MAE = The mean absolute error

ISSN: 1471-2261