Dense attentional network for pancreas segmentation in abdominal CT scans

Deep neural networks have been widely used in medical image segmentation and they can achieve good results in segmentation of some big organs. However, for some small organs, such as the pancreas in 3D CT images, the segmentation results are usually not satisfactory due to the low proportion. In this paper, we present a novel network --- Dense Attentional Network (DA-Net), to improve the pancreas segmentation in abdominal CT scans. In DA-Net, dense connection is used to combine low-level features of encoder with corresponding features of decoder, which can help to improve the utilization of feature maps (FMs). In addition, a new module for recombination and recalibration of feature maps (RRFM) and a new attentional mechanism --- deep attentional features (DAF), are used, which can excite the most discriminating features. Pancreas is segmented from 3D CT images by a coarse-to-fine mode, in which pancreas is firstly located from 3D CT images by a coarse segmentation network, and then pancreas is further finely segmented by the DA-Net. We evaluate the proposed method with 129 CT images from NIH pancreas dataset and BTCV segmentation challenge, and compare it with several mainstream segmentation networks. Comparing with these networks, our DA-Net has the higher mean DSC of 81.39%. This shows the effectiveness and advantage of the proposed method.