U-Net: Convolutional Networks for Biomedical Image Segmentation

Symmetric encoder-decoder fully-convolutional network for dense pixel-wise biomedical image segmentation — contracting path with channel-doubling 3×3 convs and max-pool downsampling, expansive path with up-convs and skip concatenation of cropped encoder features, trained from scratch on tens of images via heavy elastic-deformation augmentation and a distance-weighted cross-entropy loss that learns inter-instance separation borders.