基于2-3D混合卷积神经网络的胰腺CT图像自动分割算法研究  

Automatic segmentation algorithm of pancreatic CT images based on 2-3D mixed convolutional neural network

作　　者：宁民权 文静[1] 黄意 王翊[1] NING Minquan;WEN Jing;HUANG Yi;WANG Yi(College of Computer Science,Chongqing University,Chongqing,400044,China)

机构地区：[1]重庆大学计算机学院,重庆400044

出　　处：《第三军医大学学报》2021年第18期1729-1734,共6页Journal of Third Military Medical University

基　　金：国家自然科学基金面上项目(61672120);中央高校基本业务费(2020CDCGJSJ043,2020CDCGSL054,2019CDYGZD004)。

摘　　要：目的探讨CT图像中胰腺组织自动分割的深度学习算法。方法以2D胰腺分割网络和3D胰腺分割网络为基础,利用2D分割网络提取的判别性语义信息,3D分割网络提取出三维空间信息,最后将判别性的语义信息和三维空间信息进行融合实现胰腺分割。结果本研究在NIH数据集上的进行了验证,平均DICE系数达到83.2%(最大值90.65%,最小值67.04%)。分别超出了2D基准方法和3D基准方法0.77%和1.38%。其胰腺分割结果细节准确,边缘平滑。2-3D方法分割出的胰腺组织相比于2D方法和3D方法分割出的胰腺,在轮廓形态上与对应的手工标注图吻合度更高,更准确。相比于2D方法,2-3D方法有效综合了三维空间信息,对胰腺边缘位置进行了补足,避免了在胰腺边缘出现的漏分割情况。相比于3D方法,2-3D方法有效地利用了2D方法提供的语义信息,规避了在胰腺边缘出现的错误分割的情况。结论该方法融合了2D分割网络和3D分割网络的优点,能够有效地应用于对胰腺的准确分割。Objective To investigate the deep learning algorithm for automatic segmentation of pancreatic tissue from CT images. Methods Based on both 2 D and 3 D pancreatic segmentation networks, which extracted discriminative semantic information and three-dimensional spatial information from 82 CT images of pancreas samples(taken from NIH database) respectively, a novel hybrid network was designed for automatic segmentation of pancreatic tissues. Results The proposed method was validated using NIH datasets, and the results of pancreas segmentation demonstrated accurately in detail and smooth in edge. The average Dice coefficient reached 83.2%, which exceeded the 2 D and 3 D benchmark methods by 0.77% and 1.38%, respectively. Conclusion This method combines both advantages of 2 D and 3 D segmentation networks, and thus can be effectively applied to the accurate segmentation of pancreas, providing a quantitative reference for subsequent clinical diagnosis and treatment.

关 键 词：胰腺分割 深度学习 计算机辅助诊断 2-3D网络融合 

分 类 号：R319[医药卫生—基础医学] R322.491