Pancreas segmentation in computed tomography imaging has been historicallydifficult for automated methods because of the large shape and size variationsbetween patients. In this work, we describe a custom-build 3D fullyconvolutional network (FCN) that can process a 3D image including the wholepancreas and produce an automatic segmentation. We investigate two variationsof the 3D FCN architecture; one with concatenation and one with summation skipconnections to the decoder part of the network. We evaluate our methods on adataset from a clinical trial with gastric cancer patients, including 147contrast enhanced abdominal CT scans acquired in the portal venous phase. Usingthe summation architecture, we achieve an average Dice score of 89.7 $\pm$ 3.8(range [79.8, 94.8]) % in testing, achieving the new state-of-the-artperformance in pancreas segmentation on this dataset.
展开▼
