Figure 1.

(top) Examples of different biopsy tissues with glands classified as Gleason grade 3 or 4. (bottom) Deep convolutional neural network architecture: a biopsy patch image is fed into the network and the output after a softmax normalization is composed of 7 segmentation maps 32 times smaller than the input. The network is composed of 6 consecutive conv blocks. A conv block consists of 2 consecutive convolutions of input features (f feature maps of height h and width w) with a squeeze-and-excitation block and a residual connection. Downsampling is done using strided convolutions with a kernel 2x2, after which batch normalization (BN) and an activation rectifier linear unit (ReLU) are applied.