(Full PyTorch Code Reference: https://github.com/milesial/Pytorch-UNet)

1 Introduction

• Pooling layers → replaced by upsampling operators - These layers increase the resolution of the output
• High resolution features from the contracting path are combined with the upsampled output

Network does not have any fully connected layers and only uses the valid part of each convolution

→ This strategy allows the seamless segmentation of arbitrarily large images by an overlap-tile strategy (so there are no pad maybe)

Very little training data available → use excessive data augmentation (by applying elastic deformations to the available training images)

Propose the use of a weighted loss: the separating background labels b/w touching cells → obtain a large weight in the loss function

2 Network Architecture

Consists of a contracting path (left side) and an expansive path (right side)

• At each downsampling (2x2 max pooling with stride 2) step → double the number of feature channels
• At each upsampling (2x2 up-conv) step → half the number of feature channels
• After upsampling → concatenation with the correspondingly cropped feature map from the contracting path (cropping is necessary due to the loss of border pixels in every convolution)