From fat droplets to floating forests: cross-domain transfer learning using a PatchGAN-based segmentation model