AG-UN-Net: U-shaped and N-shaped network with attention gate for liver segmentation

U-shaped network is the most widely used backbone architecture in the field of biomedical image segmentation in recent years. Meanwhile, it is widely used in segmentation of liver from computed tomography, since these volumes have heterogeneous and diffusive shapes. However, there are considerable performances drop in the case of detecting small region, finer details of surfaces and edges. We analyze this issue and address it by adding N-Net to backbone, where N-shaped network maps data to higher dimensions to segment small region and blurred boundaries. Meanwhile, we replace skip connection with attention gate in U-shaped net. The soft-attention can highlight specific target regions and avoid the model to repeatedly extract similar low-level features. The proposed method is validated on the public Liver Tumor Segmentation Challenge dataset. Experimental results demonstrate the superiority of the proposed method compared with the other advanced methods, where the dice scores of AG-UNet and AG-UN-Net are 96.7% and 97.8%, respectively.