基于EM路由算法的医学图像分割UCaps网络

UCaps Network Based on EM-Routing Algorithm for Medical Image Segmentation

传统的医学图像分割网络存在分割精度低、图像信息易丢失、分割轮廓不清晰等问题。为提高医学图像分割准确率,提出一种结合胶囊网络与U-Net的多标签图像分割网络UCaps。以U-Net网络为架构,基于胶囊网络原理设计适用于胶囊网络的上采样算法,通过结合高斯混合模型作为聚类算法的EM路由算法聚合底层特征对高层特征的推导过程,使高层特征包含底层特征信息,同时底层特征间的位置、姿态等信息具有统一性。实验结果表明,相比U-Net、SegCaps、MaVec-Caps网络,UCaps网络的平均分割准确率为93.21%,其中左肺分割准确率达到98.24%,具有较高的图像分割准确率和较快的收敛速度。

The existing medical image segmentation networks are limited in segmentation accuracy,and often lose image information as well as produce vague segmentation boundary.In this paper,we propose a multi-label image segmentation network named UCaps,which combines capsule networks with U-Net.Taking U-Net as the basic structure,we designs an up-sampling capsule algorithm based on the principle of capsule network.Combining EM-Routing algorithm with Gaussian Mixture Model(GMM)is used to cluster the probability inferences made by child capsules for parent capsule,so the parent capsule can keep both the feature information of child capsules and the consistency of location,posture and other information between child capsules.The experimental results show that compared with UNet,SegCaps and MaVec-Caps network,the average segmentation accuracy of UCaps network is 93.21%,and the accuracy of left lung segmentation reaches 98.24%.The proposed network significantly improves the segmentation accuracy and convergence speed.