Paper Review 14: U-Net: Convolutional Networks for Biomedical Image Segmentation

Summary

1. Contracting path to capture context
2. Expanding path that enables precise localization

Loss Function

pixel-wise soft-max over final feature map combined with cross entropy loss function

\[E=\sum_{x\inΩ}w(x)\log(p_{l(x)}(x))\\ w(x)=\text{weight map for giving more importance to some pixels}\\ l:Ω→\{1,...,K\} = \text{true label of each pixel}\]

Training

Pre-compute the weight map for each ground truth segmentation

1. to compensate the different frequency of pixels from certain class in the training data set
2. force network to learn small separation borders

\[w(x)=w_c(x)+w_0\exp \left(-\frac{(d_1(x)+d_2(x))^2}{2σ^2}\right)\]

• w_c = weight map to balance class frequencies
• d1 = distance to the border of nearest cell
• d2 = distance to the border of second nearest cell