Liver Vessel

 Segmentation of liver vessels is of great importance to disease diagnosis and surgical planning. However, the segmentation performance of existing models are far from satisfactory due to the lack of elaborated dataset and problem-specific sophisticated methods. Although many deep learning models have shown great capability in addressing various segmentation tasks, a large number of well labeled data is required. To this end, we build a high-quality and large-scale liver vessel dataset, which contains 325 CT volumes and about 10 thousands slices. Based on this, an attention guided bidirectional-scaling deep learning-based model is conceived. Experimental results show that the proposed method makes a big jump on vessel segmentation, which achieves an increment of dice similarity score by 11% and 15% obtained from the newly constructed dataset than that of the well known dataset 3D-IRCADb and MSD, respectively. Obviously, the newly constructed dataset is promising to the formulation of more accurate vessel segmentation models.