Fig. 4From: Comparison of automated segmentation techniques for magnetic resonance images of the prostateArchitecture of the GAN discriminator. The ground truth and generated masks are first concatenated, then passed through consecutive strided convolution blocks (blue) and one regular convolution block. All blocks are non-residual Convolution blocks with \(4\times 4\times 4\) kernels followed by batch norm (apart from the first strided block) and a PReLU activation. The final dense layer uses a linear activationBack to article page