No abstract

Liver image segmentation is an attractive topic in the diagnosis and surgical planning of liver diseases. Although deep learning methods have significantly advanced liver segmentation, existing frameworks fail to clearly determine liver boundaries, especially in medical images where various organs have similar grey levels. In this paper, the authors design a multi-scale dense residual network (MDR-Net) for liver segmentation, which consists of two blocks: a liver segmentation network and an edge-aware network. In the segmentation network, the authors introduce a multi-scale residual pooling module combining channel attention (CA) mechanism and depth-wise separable convolution to accommodate liver scale variation. Furthermore, the authors employ an edge-aware loss network to refine edge information and enhance feature representation, which is beneficial to guide the network to iterate towards the ground truth. The authors' method achieves the best visualization results in qualitative evaluation. In addition, the authors' method achieves 96.189% on 3D-IRCADb and 96.889% on the CHAOS dataset in quantitative evaluation with respect to the dice index.