Automatic segmentation of liver CT images based on dense pyramid feature network

Objective To discuss the performance of fully automatic liver segmentation in multi-phase abdominal enhanced CT images by dense pyramidal feature network. Methods We collected the DICOM images of abdominal enhanced CT of 20 patients admitted to the first medical center of Chinese PLA General Hospital from 2015 to 2018. There were 15 males and 5 females aged over 30 years. The data were analyzed by Python and Tensorflow framework via constructing a dense pyramid feature network for automatic liver segmentation, and its performance was compared with the widely adopted U-Net structure. Results In the experiment, the DICE coefficient of liver segmentation in the arterial phase, venous phase, and delay phase by dense pyramid feature network was 95.97%, 96.22%, and 96.16%, respectively, higher than 95.59%, 95.85%, and 95.56% by U-Net structure. Conclusion The results show that dense pyramid feature network is superior to U-Net segmentation network for multi-phase liver segmentation in different phases.