Fig. 2

Convolutional variational autoencoder structure. The input is a 6-channel 128 × 128 pixel image representing the three-dimensional coordinates of epicardium and endocardium. Each encoder block (EncBlock) features a 2D convolutional layer (kernel size of 3, stride 1), batch normalization and a ReLu activation function. Each decoder block  (DecBlock features an upsampling bilinear layer of factor 2, a 2D convolutional layer (kernel size of 3, stride 1) and a ReLu activation function, which is not used in the last block. A ReLu activation function is also used for the linear layers. The red box represents the latent space variables