Fig. 10

Schematic representation of the network training of the used U-net architecture to segment live cell images of endothelial cells in wound healing experiments. The network is composed of 3 down-sampling convolutional steps and 3 up-sampling stages of de-convolutional layers, the dimensions and number of feature layers of the steps is noted below the blocks (x-dimension, y-dimension, feature layers). The coloring of the blocks stands for: Input Image (blue), convolutional layer (orange), de-convolutional layer (apricot), max-pooling (purple), drop out (cyan), rectifier layer – ReLU (red), classification output (green). The loss during training is computed due to Tversky-loss function. The final segmentation results are pixels that are classified in cell, border or background