deep-learningCNNsegmentationmedical

U-Net + ResNet : The Importance of Skip Connections in Biomedical Image Segmentation

Reviewed on May 8, 2017 by Pierre-Marc Jodoin • https://arxiv.org/pdf/1608.04117.pdf

Reference : Michal Drozdzal, Eugene Vorontsov, Gabriel Chartrand, Samuel Kadoury, and Chris Pal. The Importance of Skip Connections in Biomedical Image Segmentation. MICCAI Workshop DLMIA (2016).

This paper combines the U-Net with ResNet. Their main contribution is the use of “bottleneck block” and “simple block” layers which all contain dropout.

The authors show that the use of local skip connections allows to have similar results than state-of-the-art methods with fewer parameters, without any post-processing, and without any class weighting. They tested their method on the EM (Electron Microscopy) ISBI 2012 challenge dataset.