Extended Res-UNet with Hierarchical Inner-Modules for Liver Tumor Segmentation from CT Volumes
Jiayin Shi, Sei‐ichiro Kamata · 2022
Liver cancer is the second most common cause of cancer death and the sixth most frequent cancer in the world. Multi-Detector Computed Tomography (MDCT) images are now widely used to manifest liver tumors’ shape and volume, but it is hard for a doctor to realize such information at once. To support automatic MDCT analyzing work, we propose an extended residual U-Net with the novel hierarchical inner-module (HIM) and skip-connected hierarchical inner-modules (SHIMs). The HIM and SHIMs realize fine-grained feature extraction by separating the feature map into groups by channels, and using a set of small inner filter groups to extract detailed features from these groups. The inner filter group made up of one convolutional layer and one attention layer are connected hierarchically to redistribute the model’s attention on different feature groups. We evaluate the proposed method using 3DIRCADb dataset including 22 CT volumes where 16 volumes have tumors in the liver. The Dice values of liver and tumor segmentation are 0.931 and 0.792. The tumor segmentation result is better than the state-of-the-art method, showing strong ability in tumor feature extraction.