基于集成卷积神经网络的肺部肿瘤计算机辅助诊断模型

发布时间：2018-03-20 21:04

本文选题：集成卷积神经网络　切入点：肺部肿瘤　出处：《生物医学工程学杂志》2017年04期 　论文类型：期刊论文

【摘要】：将卷积神经网络(CNN)用于肺部肿瘤正电子发射计算机断层扫描(PET)/电子计算机断层扫描(CT)计算机辅助诊断,不仅可以提供精确的定量分析以弥补人眼惰性及对灰阶不敏感的缺陷,也能辅助医生准确诊疗。本文首先采用参数迁移的方法初始构建三个CNN(CT-CNN、PET-CNN、PET/CT-CNN)分别用于肺部肿瘤CT、PET、PET/CT的识别;然后以CT-CNN为例探讨迭代次数、批次大小和输入图像大小对识别率和训练时间的影响,从而选择合适的模型参数训练单一CNN;最后集成三个单一CNN,采用"相对多数投票法"完成肺部肿瘤PET/CT计算机辅助诊断,进而对比集成CNN与单个CNN的性能。实验结果表明集成CNN模型比单一CNN模型对于肺部肿瘤计算机辅助诊断的性能更优。
[Abstract]:The application of convolutional neural network (CNN) to the computerized diagnosis of pulmonary tumors by positron emission computed tomography (PET) / computerized tomography (CT) can not only provide accurate quantitative analysis to remedy the defects of human eye inertia and insensitivity to gray scale. The method of parameter migration was used to construct three CNNs CT-CNNP PET / CT-CNNNs respectively for the identification of pulmonary tumors, and then CT-CNN was taken as an example to discuss the iterations. The effects of batch size and input image size on the recognition rate and training time, so as to select the appropriate model parameters to train the single CNNs, finally integrate the three single CNNs and complete the PET/CT computer-aided diagnosis of lung tumors by "relative majority voting method". The experimental results show that the performance of integrated CNN model is better than that of single CNN model for computer-aided diagnosis of lung tumors.
【作者单位】：宁夏医科大学公共卫生与管理学院;宁夏医科大学理学院;宁夏医科大学总医院核医学科;
【基金】：国家自然科学基金(81160183,61561040) 宁夏自然科学基金(NZ16067) 宁夏高教项目(NGY2016084)
【分类号】：R734.2;TP183;TP391.7

【相似文献】