Ronneberger Olaf Fischer Philipp Brox Thomas

Abstract

There is large consent that successful training of deep networks requiresmany thousand annotated training samples. In this paper, we present a networkand training strategy that relies on the strong use of data augmentation to usethe available annotated samples more efficiently. The architecture consists ofa contracting path to capture context and a symmetric expanding path thatenables precise localization. We show that such a network can be trainedend-to-end from very few images and outperforms the prior best method (asliding-window convolutional network) on the ISBI challenge for segmentation ofneuronal structures in electron microscopic stacks. Using the same networktrained on transmitted light microscopy images (phase contrast and DIC) we wonthe ISBI cell tracking challenge 2015 in these categories by a large margin.Moreover, the network is fast. Segmentation of a 512x512 image takes less thana second on a recent GPU. The full implementation (based on Caffe) and thetrained networks are available athttp://lmb.informatik.uni-freiburg.de/people/ronneber/u-net .