3D CD-LinkNet with Cascaded Dilated Convolution for Liver Vessel Segmentation

Automatic segmentation of liver blood vessels from CT images is very helpful for clinical surgery. Due to the small volume, variable shape and low contrast of CT images, it is difficult to segment blood vessels in abdominal CT images. To address these issues, this paper attempted to segment liver blood vessels in three steps. The first step is pre-processing to improve the contrast of the CT images. The second step is to segment liver blood vessels through a newly proposed 3D segmentation network with dilated convolutions. The third step is to perform post-processing on the segmentation results. Segmenting liver blood vessels in 3D views may result in better segmentation results due to the increased volume of blood vessels. Dilated convolutions are cascaded to extract features from a wider field of view, thereby solving the issue of small volume and variable shape. This method has been tested on a public dataset 3Dircadb-01 and compared with several existing methods. Experimental results show, the proposed method achieved the highest segmentation accuracy, with Dice score 0.6596. The method proposed in this paper is limited to blood vessel segmentation in the liver region and can be used in conjunction with appropriate liver segmentation methods.