U-Net is a convolutional neural network architecture designed for dense pixel-level prediction, originally developed for biomedical image segmentation at the University of Freiburg in 2015. Its defining feature is a symmetric encoder-decoder structure with skip connections that pass high-resolution feature maps from the encoder directly to the corresponding decoder stage. This lets the network combine deep semantic understanding with fine spatial detail, producing sharp segmentation boundaries.

The encoder follows a standard contraction path: repeated convolutions and max-pooling layers progressively reduce spatial resolution while increasing feature depth. The decoder mirrors this with upsampling (transposed convolutions or bilinear interpolation) and convolution layers that gradually restore the original resolution. Skip connections concatenate encoder features at each level with the upsampled decoder features, giving the network access to both local texture and global context at every scale.

U-Net became the go-to architecture for medical image segmentation because it works well with very small training sets, a common constraint in clinical settings where labeled data is scarce and expensive. Variants like U-Net++, Attention U-Net, and nnU-Net have extended the original design with nested skip connections, attention gates, and automated architecture search. Beyond medical imaging, U-Net and its variants are used for satellite image segmentation, industrial defect detection, and as core components inside diffusion models for image generation.