Pancreas Segmentation Based on an Optimized Coarse-to-Fine Method
Chengjian Qiu, Zhe Liu, Yuqing Song, Kai Han · 2020
Our goal is to automatically segment the pancreas from the abdominal CT scans. Convolutional Neural Network (CNN) based deep learning method is easy to be confused when the background is larger than the target (pancreas) region. To solve this problem, a coarse-to-fine method has been proposed which used the predicted segmentation from the coarse stage to generate a smaller input for the fine stage to reduce the influence of background around the pancreas. However, In different stages of the coarse-to-fine approach, the accuracy isn't enough high for clinical purposes. In this paper, we want to achieve higher performance. Therefore, a variety of methods are adopted to improve the coarse-to-fine pancreas segmentation. Firstly, Elastic deformation is used to augment data, which increased data diversity. Secondly, we design a model called Aggregation-UNet by using a deep layer aggregation structure, which leverages spatial and semantic information of the data to reduce the false segmentation. Thirdly, to make the network pay more attention to the target area, we use a novel location method after the coarse stage to find a smaller region for the fine stage. Finally, we integrate the trained model to increase the reliability of the final result and the Dice accuracy. Our proposed method is evaluated on the NIH pancreas dataset which indicates the state-of-art result, while we only use the axial view of the CT scans (generally using three views).