3D Segment and Pickup Framework for Pancreas Segmentation

Abstract

Locate and segment (LAS) framework is an effective method for segmenting pancreas from abdominal CT. Coarse-to-fine is the most widely used LAS framework which has achieved excellent pancreatic segmentation results collaborated with many network architectures. However, inaccurate location of the region of pancreas reduces performance of LAS methods. To solve these problems, we propose the segment and pickup (SAP) framework, which uses manual annotation to directly calculate the ROI of pancreas during training and trains a neural network to segment the pancreas in the ROI. In the testing process, we first use the well-trained segmentation network to segment the pancreas from the whole CT scan, then use the region growing method to pick up the final segmentation results from the noise. We used ResNet combined with the SAP framework to conduct experiments on the NIH data set, and achieved 86.96 DSC scores, proving that our SAP framework performs better than the regular LAS framework on pancreas segmentation.