U-Net: Convolutional Networks for Biomedical Image Segmentation

Inside the paper

The problem

Biomedical images often need a label for every pixel, but annotated examples are scarce. Earlier patch-by-patch networks repeated work across overlapping regions and had to trade local detail against a wider view.

The idea

U-Net joins a shrinking path that gathers context to an expanding path that reconstructs a segmentation. Training uses transformed examples and gives extra attention to the narrow gaps between touching cells.

How it works

• Repeated convolutions and pooling reduce spatial size while increasing the number of feature channels.
• Upsampling restores resolution, and cropped features from the matching earlier layer supply fine spatial detail to each stage.
• Elastic deformations create varied training examples; a weighted pixel loss emphasizes separating borders as well as class balance.
• For large images, overlapping tiles and mirrored image edges provide the context needed to predict the output near boundaries.

What it showed

On the electron microscopy challenge, the network achieved a lower warping error than the listed sliding-window network. It also produced higher intersection-over-union scores than the second-best entries on both reported cell datasets.

Read it for

Begin with Figure 1 and Section 2 to trace the two paths, then read Section 3 and Figure 3 for the training choices. Section 4 and Tables 1 and 2 show what was tested; the implementation details of tiling can be skimmed on a first pass.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.