CR-UNet: Context-rich UNet for liver segmentation from CT volumes

Automatic segmentation of liver from CT volumes is essential for diagnosis and treatment of liver cancer. In order to achieve accurate segmentation of liver, we proposed a novel 3D residual UNet with context collection module (CCM). The proposed CCM can capture the long-range correlations between all pixels, thereby producing rich and dense contextual information with full-image dependency. We applied CCM at the end of contraction path of UNet to ensure the features fed into expanding path rich in contextual information. Therefore, our model is more accurate and reliable for liver edge segmentation. After extensive experiments, our model finally achieved a state-of-the-art result as Dice per case score of 0.961 and Dice global score of 0.963.