【作者】 王卓；

【导师】 王卫兵；

【作者基本信息】 哈尔滨理工大学 ， 软件工程， 2020， 硕士

【摘要】 肺癌是最常见的癌症和癌症死亡的主要原因,且是一种预后较差的恶性疾病,患者的5年平均生存率不到20%。因此,早期发现肺部病变以提高治疗几率并增加患者生存率非常重要。然而,由于良性和恶性肺癌结节之间的细微差别,即使对于人类专家来说,肺癌诊断也是一项艰巨的任务。因此需要发展计算机辅助诊断系统,以协助医生对肺癌结节进行诊断,提供其诊断准确率。本文以肺癌CT图像为研究基础,着重研究肺癌图像分类,以帮助医生对早期肺癌患者进行诊断治疗,提高治愈几率。论文首先介绍了国内外深度学习的研究现状和肺癌图像分类方法的研究现状,然后对肺癌图像分类典型CNN模型进行分析,对各主流CNN模型分析其特点,对Alex Net、ZFNet、Google Net、VGGNet进行性能比较。实验表明,基于原始图像,VGG16相较于其他网络结构模型具有更高的分类准确率。在此基础上,基于VGG16卷积神经网络进行肺癌图像分类,选择公共肺癌数据集LIDC-IDRI,对肺癌图像进行预处理,通过实验选取激活函数和设置模型参数,选取准确率、敏感性、特异性和ROC曲线为指标与其他论文方法作对比分析。实验表明基于VGG16的肺癌图像分类准确率和敏感性均有小幅度提升,但仍存在训练时间长等问题。针对VGG16网络层数深,参数过多的问题,提出一种优化的网络结构,以减少模型参数,加快网络收敛速度,减少模型的过拟合问题。通过LIDC-IDRI数据集上进行的训练和测试,结果表明所提出的优化的卷积神经网络可以从三维图像中提取肺癌结节信息,并有效地将其信息用于肺癌图像分类。更多还原

【Abstract】 Lung cancer is the most common cancer and the main cause of cancer death,and is a malignant disease with poor prognosis.The average 5-year survival rate of patients is less than 20%.Therefore,it is important to detect lung lesions early to increase the chance of treatment and increase the patient survival.However,due to the subtle differences between benign and malignant lung nodules,even for human experts,the diagnosis of lung cancer is a difficult task.Therefore,a computer-aided diagnosis system needs to be developed to assist doctors in diagnosing lung cancer nodules and provide their diagnostic accuracy.Based on CT images of lung cancer,this paper focuses on the classification of lung cancer images,so as to help doctors diagnose and treat patients with early lung cancer and improve the chances of cure.Firstly,this paper introduces the research status of deep learning at home and abroad and the research status of lung cancer image classification methods.Then,the typical CNN model of lung cancer image classification was analyzed,the characteristics of each mainstream CNN model were analyzed,and the performance of Alex Net,ZFNet,Google Net and VGGNet were compared.Experiments showed that,based on the original image,VGG16 has a higher classification accuracy than other network structure models.On this basis,lung cancer images were classified based on VGG16 convolutional neural network,public lung cancer data set LIDC-IDRI was selected,lung cancer images were preprocessed,activation functions were selected and model parameters were set through experiments,and accuracy,sensitivity,specificity and ROC curves were selected as indicators for comparative analysis with other paper methods.The experiment showed that the accuracy and sensitivity of lung cancer image classification based on VGG16 had a small increase,but there were still problems such as long training time.Aiming at the problem of too many parameters and too many layers in VGG16 network,an optimized network structure was proposed toreduce the model parameters,accelerate the network convergence speed and reduce the over-fitting problem of the model.Through the training and testing on the LIDC-IDRI data set,the results show that the proposed optimized convolutional neural network can extract lung cancer nodule information from 3d images and effectively use the information to classify lung cancer images.更多还原