High-Resolution Recurrent Gated Fusion Network for 3D Pancreas Segmentation

Pancreas segmentation has been challenging due to its large variations in size, shape, localization, and indistinguishable boundary. Current mainstream pancreas segmentation methods are based on the deep encoder-decoder structure, which recover high-resolution representations from encoded low-resolution representations to generate pancreas voxel masks. However, the details of the pancreas are easily lost in the encoding stage. In this paper, we propose a 3D high-resolution network (3D HRNet) to extract pancreas features, which maintains high-resolution representations throughout the whole process. We use a novel recurrent gated fusion (RGF) head to fuse high-resolution features and generate pancreas voxel masks. To reduce variable background interference, we crop the pancreas area from abdominal CT images for segmentation with a pancreas localization network. We evaluate the above propsed method on the public NIH and MSD pancreas segmentation datasets, and experiments show a competitive result with a mean Dice-Srensen Coefficient (DSC) of 85.82±4.01% on NIH and 84.22±5.91% on MSD, respectively. The lowest variance and the highest mean DSC reveal the stability of our method among current methods and its potential in the clinical setting.