基于深度学习和模型集成的肺结节分割方法  

Pulmonary Nodule Segmentation Method Based on Deep Learning and Model Integration

作　　者：朱诗生[1] 王慧娟 李淳鑫 ZHU Shi-sheng;WANG Hui-juan;LI Chun-xin(Department of Computer Science,Shantou University,Shantou 515063,China)

机构地区：[1]汕头大学计算机系,广东汕头515063

出　　处：《计算机技术与发展》2023年第2期208-213,共6页Computer Technology and Development

基　　金：广东省科技计划项目(20140401)。

摘　　要：针对CT图像中肺结节所占的比率比较小、特征复杂及分割精准度不高的难题,提出了一种基于深度学习和模型集成的肺结节分割方法。该方法在数据采样上,为解决胸部CT图像中存在的类别不平衡问题和避免模型对图像中肺结节位置的过度学习,提出了一种新的随机方向采样方式。首先,将采样图截成64*64的大小;然后,在对CT图像进行分割预测时采用步长为32的遍历预测叠加方式,来避免肺结节被遗漏的现象,以提升模型性能;在分割结果上,提出在卷积网络后连接条件随机场,通过结合肺结节相邻像素点的信息来优化分割的结果;在此基础上,创新性地将多种深度学习模型(U-Net、LinkNet和SegNet)的肺结节分割结果进行集成,从而进一步提升肺结节分割的精准度。在LIDC-IDRI肺结节公开数据库上的实验验证结果表明,该方法可以更有效地提高肺结节分割的精准度,更有助于提升医生对肺癌的诊治水平。In view of the problems that small proportion of pulmonary nodules in CT images,complex features and low segmentation accuracy,a pulmonary nodule segmentation method based on deep learning and model integration is proposed.In terms of data sampling,a new random direction sampling method is proposed by considering the class imbalance in chest CT images and avoid over-learning the location of pulmonary nodule by the model.Firstly,the sampling figure is cut into a size of 64*64.Then,in the segmentation prediction of CT images,the method of traversal prediction superposition with a step size of 32 is adopted to avoid the phenomenon of missing pulmonary nodules,so as to improve model performance.In terms of segmentation results,we propose to optimize the segmentation results by means of the fully-connected conditional random field.On this basis,the pulmonary nodule segmentation results of multiple deep learning models(U-Net,LinkNet and SegNet)are innovatively integrated,so as to further improve the accuracy of pulmonary nodule segmentation.The experimental verification results of LIDC-IDRI pulmonary nodule open database show that the proposed method can improve the accuracy of pulmonary nodule segmentation more effectively,which is more conductive to improving doctors’diagnosis and treatment level of lung cancer.

关 键 词：深度学习 肺结节 分割 全连接条件随机场 集成学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] R814.42[自动化与计算机技术—计算机科学与技术]