Pancreas Co-segmentation based on dynamic ROI extraction and VGGU-Net

Abstract

Pancreas segmentation is one of the most challenging tasks in medical image segmentation for its anatomical variability, large individual differences, irregular shape, small volume and complex surroundings. Many methods can achieve more accurate segmentation results on abdominal organs other than the pancreas. To overcome this problem, this paper proposes a novel segmentation method ROI-VGGU-Net (Region of Image-Vision Geometry Group U-shaped Net) that can segment pancreases more accurately from a dynamical extracted region-of-interest (ROI) by our proposed deep learning model VGGU-Net, where the ROI is obtained by combining various location information of other surrounding organs (such as liver, spleen and kidney). Specifically, we first obtain the location of other organs around a pancreas through our proposed VGGU-Net. And then, we compute the center points of these located surrounding organs in every CT slice sequentially to form different ROIs of the pancreas through these center points. Finally, an accurate pancreas segmentation can be achieved through VGGU-Net according to the acquired ROIs. Our various experiments demonstrate the effectiveness of ROI-VGGU-Net.