查询结果:   张丽,强彦,张小龙,王三虎.基于同步深度监督的多尺度肺结节分类[J].计算机应用与软件,2019,36(9):214 - 219.
中文标题
基于同步深度监督的多尺度肺结节分类
发表栏目
图像处理与应用
摘要点击数
387
英文标题
CLASSIFICATION OF MULTI-SCALE LUNG NODULES BASED ON SYNCHRONIZED DEEP SUPERVISION
作 者
张丽 强彦 张小龙 王三虎 Zhang Li Qiang Yan Zhang Xiaolong Wang Sanhu
作者单位
太原理工大学计算机科学与技术学院 山西 太原 030024 宾夕法尼亚州立大学信息科学与技术学院 宾西法尼亚州 尤尼弗西蒂帕克 16802 吕梁学院计算机科学与技术系 山西 吕梁 033000   
英文单位
College of Computer Science and Technology, Taiyuan University of Technology, Taiyuan 030024, Shanxi, China College of Information Science and Technology, Pennsylvania State University, University Park 16802, Pennsylvania, USA Department of Computer Science and Technology,Lvliang University, Lvliang 033000, Shanxi, China   
关键词
同步深度监督 多尺度 卷积神经网络 特征提取
Keywords
Synchronized deep supervision Multi-scale Convolutional neural network Feature extraction
基金项目
国家自然科学基金项目(61572344);虚拟现实技术与系统国家重点实验室开放基金项目(VRLAB2018A08);山西省回国留学人员科研项目(2016-038)
作者资料
张丽,硕士生,主研领域:医学图像处理,深度学习。强彦,教授。张小龙,教授。王三虎,教授。 。
文章摘要
针对在肺结节分类中容易产生过拟合的问题,提出一种基于同步深度监督的多尺度肺结节分类方法。解决梯度消失问题,最小化分类错误并实现同一框架中同步训练多尺度肺结节图像,提高肺结节分类精确度。改进经典的AlexNet网络,使其更适合肺结节图像分类;将同步深度监督(SDS)策略纳入到AlexNet架构中,向隐藏层提供集成的同步监督;通过多尺度空间金字塔策略提取多尺度肺结节图像特征。实验结果表明,该方法的准确性达到93.68%,且在准确性、敏感度、特异度、ROC曲线下面积值上均优于其他分类方法。
Abstract
To solve the problem of over-fitting in the classification of lung nodules, We proposed a classification method of multi-scale lung nodules based on synchronized deep supervision. It solved the problem of gradient disappearance, minimized classification errors, and achieved simultaneous training of multi-scale pulmonary nodule images in the same framework. It improved the accuracy of pulmonary nodule classification. The classic AlexNet network was improved to make it more suitable for the classification of images of lung nodules. The synchronized deep supervision(SDS) strategy was integrated into AlexNet architecture to provide integrated synchronized supervision to the hidden layers. And the multi-scale spatial pyramid strategy was used to extract the features of multi-scale lung nodules. The experimental results show that the accuracy of this method is 93.68%. It is superior to other ones in terms of accuracy, sensitivity, specificity, and area under the ROC curve.
下载PDF全文