Region Context Aggregation Network for Multi-organ Segmentation on Abdominal CT

Pointing at the problem of the automatic segmentation of multiple abdominal organs on CT, we propose a coarse-to-fine based 3D network, named as RCANet, which could effectively refine the coarse segmentation by an end-to-end learning strategy through exploring more contextual information. Our network consists of several simple but useful modules which are helpful to represent the relation between voxels and object regions more effectively. First, we learn a 3D coarse segmentation map through a classical 3D UNet. Second, we use a region concentration block (RCB) to extract the global context information of each object region. Last, we augment the combination of each voxels and its affiliated region by utilizing a region aggregation module (RAM) and obtain the final segmentation result. In our paper, we demonstrate the advantages of RCANet on TCIA public dataset with the improvement on some small organs and on average compared with some advanced methods.