基于级联式深度网络模型的胃及胰腺自动分割研究

目的:旨在研究自主创新设计的级联式深度卷积神经网络VB-Net在胃和胰腺上的自动分割精度及效率。方法:回顾分析150例胰腺癌患者临床资料，随机选取132例非增强CT数据和其中116例胰腺期增强CT以及结构数据进行胃及胰腺的分割模型训练。对剩余18例患者的非增强CT和胰腺期增强CT给予模型测试，使用戴斯相似性系数量化分析模型的分割精度，同时评估其分割效率。结果:基于非增强CT的胃、胰腺的自动分割平均DSC值分别为87.93%、80.05%；基于胰腺期增强CT的胃、胰腺自动分割平均DSC值分别为89.71%、84.79%。胃及胰腺的自动分割平均时间为1.22、0.84 s，手动分割平均时间为158.70、115.52 s。结论:基于VB-Net的胃及胰腺自动分割模型测试结果较为准确，且极大提高了器官分割的效率。

Automatic segmentation of the stomach and pancreas using cascaded deep convolutional neural network

Abstract: Objective To analyze the accuracy and efficiency of self-developed cascaded deep convolutional neural network (VB-Net) in the automatic segmentation of the stomach and pancreas. Methods The clinical data of 150 patients with pancreatic cancer were analyzed retrospectively. The non-enhanced CT data of 132 cases and the pancreas contrast-enhanced CT and structure data of 116 cases among the 132 cases were randomly chosen as training set for training the model for segmentation of the stomach and pancreas. The non-enhanced CT data and pancreas contrast-enhanced CT data of the other 18 cases were regarded as test set for evaluating the performance of the proposed method. Finally, Dice similarity coefficient was used to quantitatively analyze segmentation accuracy and segmentation efficiency was also evaluated. Results The average Dice similarity coefficients of the automatic segmentation of the stomach and pancreas were 87.93% and 80.05% on non-enhanced CT, 89.71% and 84.79% on pancreas contrast-enhanced CT, respectively. The average automatic segmentation time of the stomach and pancreas was 1.22 s and 0.84 s, while the average manual segmentation time was 158.7 s and 115.52 s. Conclusion The VB-Net based method for the automatic segmentation of the stomach and pancreas can achieve an accurate segmentation results, and significantly improve organs segmentation efficiency.