UNet

U-Net is a convolutional neural network architecture designed for fast and precise segmentation of biomedical images. It was introduced by Olaf Ronneberger, Philipp Fischer, and Thomas Brox in 2015 at MICCAI. The design follows a U-shaped encoder–decoder structure. The contracting path (encoder) captures context through repeated blocks of two 3x3 convolutions, each followed by a rectified linear unit and a 2x2 max-pooling operation for downsampling. The expansive path (decoder) enables precise localization by upsampling of feature maps and concatenation with corresponding high-resolution features from the encoder via skip connections, followed by two 3x3 convolutions and ReLU activations. A final 1x1 convolution maps feature representations to the desired number of segmentation classes.

U-Net is typically implemented as a fully convolutional network that preserves spatial information through skip connections, allowing precise boundary delineation even with relatively small training datasets. Variants include 2D U-Net for planar images and 3D U-Net for volumetric data, as well as extended forms like U-Net++ and Attention U-Net that modify skip connections or introduce attention mechanisms.

Applications include biomedical image segmentation tasks such as cell segmentation, tissue and organ delineation in histology, MR and CT image segmentation, and other domains requiring pixel-level labeling. Training often relies on data augmentation to mitigate limited data, and loss functions such as cross-entropy or Dice loss to handle imbalanced classes. While powerful, U-Net can be computationally intensive and may require careful design choices for deep or high-resolution inputs.