U-net: Convolutional Networks for Biomedical Image Segmentation

Abstract: In this talk I will present our u-net for biomedical image segmentation. The architecture consists of an analysis path and a synthesis path with additional shortcut-connections. The network is trained end-to-end from scratch with a very low number of annotated images per application. The surprisingly simple strategy to train a network with such a low number of samples is a data augmentation with elastic deformations. Furthermore the u-net can segment arbitrarily large images with a seamless tiling strategy. For 3D images we have developed a training strategy to learn a full 3D segmentation from a few annotated slices per volume. This can be used in a semi-automated setup to create a dense segmentation from the sparse annotations on the same volume, or in a fully-automated setup, where sparse annotations speed up the training data generation significantly.

References:
O. Ronneberger, P. Fischer, and T. Brox: U-Net: Convolutional Networks for Biomedical Image Segmentation.
MICCAI 2015, available at http://arxiv.org/abs/1505.04597
Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, O. Ronneberger: 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. MICCAI 2016, available at http://arxiv.org/abs/1606.06650