U-Net: segmentation from a few images
On 18 May 2015 Olaf Ronneberger and colleagues in Freiburg described U-Net, a convolutional network whose contracting path is joined to a symmetric expanding one. With it they won the ISBI 2015 cell tracking challenge in two categories, and a 512x512 image is segmented in under a second.
Why it matters
The network learned from very few annotated images, leaning on augmentation by elastic deformation, and gave precise object boundaries. The same shape later became the backbone of diffusion models for image generation.
On the ISBI challenge for segmenting neuronal structures in electron microscopy: 30 training images of 512x512 and a warping error of 0.0003529 against 0.000420 for the best earlier submission. In the 2015 cell tracking challenge, IOU 0.9203 on PhC-U373 and 0.7756 on DIC-HeLa.