Figure 3
The network architecture of pix2pix-cGANs for translating a 3D cube with air artifacts (source) to a 3D cube without air artifacts (target). The network comprises a generator (U-Net) and a discriminator (PatchGAN), similar to the pix2pix network proposed by Isola et al. (2017 ), but with network blocks adjusted to accommodate 3D inputs and outputs. The generator receives cubes with air artifacts as input (Real A) and generates cubes which are potentially free from air artifacts as output (Gen B). Pairs of Gen B and Real A (fake pair) are compared with pairs of Real B and Real A (real pair) using the discriminator network. Over the course of prolonged training, the generator network learns to generate outputs that can deceive the discriminator network into misclassifying fake pairs as real ones. |