A comparative analysis on liver segmentation from CT images -ResNet and VGG
Jie Shu, Renxiang Wang, Zhishuai Feng · 2023
Most of the segmentation of liver in CT images used U-shaped structure. Although the segmentation results of these studies are superior to classical UNet, there is no study comparing the effects of using different encoders in U-shaped structures. In this paper, we compared the use of two classic encoder structures, ResNet and VGG, to form a U-shaped structure for liver segmentation in CT images. The U-shaped structure formed here is based on the classic network structure of UNet and LinkNet, and has been tested on the public dataset 3Dircadb. The experimental results show that using VGG19 encoder combined with Linknet network structure achieved the best liver segmentation results, with a mean Dice 0.963 and a segmentation speed of about 62ms per slice. In addition, LinkNet, which uses the ResNet18 encoder, has the fastest segmentation speed, with approximately 35ms per slice, and its mean Dice value of liver segmentation results is 0.955.