– modified and extended fullyconvolutional network (FCN)
– 2015 Winner of the cell Tracking Challenge (ISBI)
– provides good segmentation for a small training data set
– is well suited for the analysis of biomedical images with complex biochemical processes due to its contextual information

Fully convolutional (deep neural) network (FCN)

– each pixel from the output image is a result of a calculation over a corresponding patch in the input image

→ size of this patch is called: the receptive field (RF) of the network

– no dense layer is used

fcn
FCN architecture

U-Net Architecture

– U-shaped architecture
– consists of a contracting and a, symmetrically constructed, ex-
breading path

unet original
Architecture of the Fully Convolutional Network (FCN) for biomedical
Image segmentation (U-Net)

Contracting part

– captured by a large number of filters Context-information
→ classic CNN
– At each reduction step the total number of filters is doubled

Expanding part

– here the localization is done
– Halving of the feature map by upsampling with subsequent 2×2 covolution at each step
– The network has no fully connected layers
– Since the valid parts of each covolution are used, the segmentation consists only of pixels for which a context is available in the input image