Deep Neural Network Guided by Attention Mechanism for Segmentation of Liver Pathology Image

Tumor region segmentation in liver pathological images has become a significant tool of routine diagnosis and it also plays an important role in the liver cancer surgical plannings. Liver lesions are usually manually marked by clinicians on ultra-high resolution pathological images, which is awfully time-consuming and laborious. Aiming at this problem, we propose a new deep neural network for automatically segment liver cancer pathological images in this paper to assist doctors in completing the diagnosis efficiently. The existing U-Net framework was mainly used in previous researches, but not all features extracted by the encoder module are suitable for pathological image segmentation. In order to solve this problem, the network in this paper designs an attention-guided connection module to obtain accurate segmentation of liver cancer, so that the useful context features can be selected adaptively from the low-level features which are guided by the high-level features. The pyramid sampling operation is applied to the global attention module to fuse different levels of features. The results of experiments show that the method has some certain advantages over the previous methods in liver cancer pathological image segmentation datasets, and its IOU is increase by 5.4% and 2.2% compared with U-Net and PSPNet networks. Simultaneously, in order to testify the basic concept preponderance, some related studies about the ablation are carried out.