DV-Net: Accurate liver vessel segmentation via dense connection model with D-BCE loss function

Abstract Recently, liver vessel segmentation has aroused widespread interest in medical image analysis. Accurately extracting blood vessels from livers is a difficult task due to their complex vessel structures and image noises. To make a neural network better adapt to this complexity, a deeper neural network is required to fit this nonlinear transformation. In particular, accurate segmentation of small blood vessels is always a challenge, since onefold down-sampling usually causes the loss of information. In this study, we introduce a dense block structure into the V-net to construct a new Dense V-Net (DV-Net) and use data augmentation to segment liver vessels from abdominal CT volumes with a few training samples. In addition, we propose a dual-branch dense connection down-sampling strategy (DCDS) to better capture vascular features and a D-BCE loss function to maximize the utilization of image resources. The proposed DV-Net structure is more powerful in the discrimination of vessel and non-vessel areas. We extensively evaluated the proposed method on the datasets of 3Dircadb and MICCAI 2018 Medical Segmentation Decalthon (MSD) Challenge. Experimental results show that the proposed DV-Net significantly improves the average segmentation Dice score. The average Dice score and sensitivity on 3Dircadb were (75.46%) and (76.93%), respectively, which are better than those of existing methods.