Attention V-Net: A Residual U-Net with Attention Gate Block for Lung Organs At Risk Segmentation

Abstract

In this paper, we try to incorporate residual connection and Attention Gate block into medical image segmentation network. At first, we construct a 2D residual U-Net (a 2D V-Net) to incorporate residual connection for medical image segmentation. In order to incorporate Attention Gate block into the V-Net, we build up the Attention Residual Block which adds a shortcut into Attention Gate Block. The Attention Residual Block will be more adaptive than raw Attention Gate Block. We also insert the Attention Residul Block into the skip connection between the encoder and the decoder of 2D V-Net and create a new network called Attention V-Net. Then we train and evaluate the networks on the 16th CSTRO conference Lung OAR segmentation competition dataset. What's more, we find out when the mirrored OARs are segmented, the networks may mix up them together. Therefore, we use a postprocessing method to correct the result. Finally, we compare the model with the state-of-the-arts to show the superiority of the proposed network.