【作者】 李畅；

【导师】 周柚；

【作者基本信息】 吉林大学 ， 软件工程（专业学位）， 2020， 硕士

【摘要】 肺癌是一种高发病率和死亡率的癌症,是世界最常见的恶性肿瘤之一。肺癌发病率持续增长的原因与人们周围所生活的环境有关,也与人们一些不良的生活习惯有关。随着城市工业化现代化速度和农村城镇化城市化速度的加快,包含食品污染,水源污染,土壤污染以及大气的污染都呈现加剧的形式,特别是我国北方地区的大气污染更为严重,这些因素与呼吸系统疾病的发病有关,是直接影响肺癌发病的因素。但更为重要的因素是对疾病的初期筛查与诊断,许多肺癌早期患者并没有出现可判断性的不良症状是导致病情无法及时发现的主要原因,后果就是会形成一经出现临床症状就诊检查时已经发展为癌症晚期阶段的情况,此时发现确诊的癌症病例中80%以上的患者都是肺癌晚期或者局部肺癌晚期,失去了最好的治疗时机。因此,能否有效识别肺结节,对肺结节进行自动良性和恶性的分类,对于肺部疾病的早期诊断和治疗具有重要研究意义和应用价值。本文针对肺部CT图像进行肺癌疾病分类预测,提出了结合DPN(Dual Path Network,双路径网络)网络与CapsNet(胶囊网络)的肺结节良恶性预测模型--DPNCapsNet。首先,针对一般性CT图像数据进行窗口设置、归一化处理以及肺实质提取,利用腐蚀与膨胀操作清除无关组织能较好的保留结节,排除CT图像中的过多干扰以增强模型对特征的挖掘能力与分析能力。然后,通过在LIDC-IDRI数据中医师标注的结节发展程度与LUNA16数据集进行校对,LUNA16中的肺结节被分为良性与恶性标签,将CT图像坐标系转换为像素坐标系,切割并提取结节所在位置,切割大小为48×48×48的立方体,再加上随机切割若干48×48×48大小的立方体区域作为训练输入。最后,将预处理后的训练数据以ZCA白化算法增强输入DPNCapsNet模型中训练。模型主要由三部分组成:首先,立方体数据经DPN网络,其中以卷积网络、ResNet(残差网络)以及DenseNet(密集连接网络)构成的DPN块不断提取数据特征。然后,利用胶囊网络直接搜集DPN网络提取到的特征,并在初级胶囊层以卷积操作构建为固定长度的向量胶囊。最后,在输出胶囊层利用动态路由算法将特征胶囊逐步路由至良性特征类与恶性特征类,得到固定长度的分类胶囊,其中胶囊的输出向量则包含结节不同的特征信息,分类胶囊的模长代表良恶性分类的概率。本文采用LUNA16数据集为主,以LIDC-IDRI数据集为辅的方式提取训练数据和医生标注标签,对数据集进行了充分的预处理工作,减少了冗余干扰并提升了训练效率。结合双路径网络对特征提取速度快、准确度高的优点以及胶囊网络对空间姿态、方向信息敏感的特点构建肺结节分类预测模型,其能够以较少的训练数据和次数获得稳定的结果。实验结果表明,在LUNA16数据集中,针对良恶性结节分类预测效果达到91.56%,接近职业医师水平。更多还原

【Abstract】 Lung cancer is one of the most common malignant tumors in the world and has a high rate of incidence and mortality.Continuous increase of lung cancer is related to the surrounding environment as well as habits.With industrial growth and urbanization,food,water,soil and air pollution levels are aggravated.Especially in northern China,air pollution is more serious.These factors lead to increase in respiratory diseases and may cause lung cancer in the long run.The long term harmful effects of the surrounding can be prevented by early screening and diagnosis.However,many lung cancer patients do not exhibit such symptoms in early stages which leads to the cancer not being detected on time.By the time clinical symptoms or serious conditions appear and the patient visits a hospital for medical examination,the cancer usually progresses to a late stage.Studies reveal that 80% of the patients were in the late stage of lung cancer or local lung cancer by the time they were diagnosed and had lost the best opportunity for treatment.Therefore,it is of great significance and practical value for the early diagnosis and treatment of lung diseases that the pulmonary nodules be identified and classified automatically as benign or malignant.Based on the lung CT images in medical CT images for classification and prediction of lung cancer diseases,this paper proposes a Dual Path Network and Capsule Network benign and malignant prediction model of lung nodules-DPNCapsNet.First,window setting,normalization processing,and lung parenchyma extraction are performed for general CT image data.Subsequently,corrosion and dilation are used to remove irrelevant tissues and preserve nodules and excessive interference in CT images is eliminated to enhance model feature mining and analysis capabilities.Then,the lung nodules are classified into benign and malignant tags by collating the nodule development degree marked by the doctor in the LIDC-IDRI data,and the CT image coordinate system converted to the pixel coordinate system.The position of the nodule is segmented and extracted in a size of 48 × 48 × 48 cube,plus a 48 × 48 × 48 cube area around the random location given in LUNA16 as training input.Finally,the preprocessed training data is enhanced into the DPNCapsNet model for training using the ZCA whitening algorithm.The model is mainly composed of three parts.First,the cube data is passed through the DPN network.The DPN block consisting of a convolution network,ResNet(Residual Network)and DenseNet(Dense Connection Network)continuously extracts data features.Then,the features extracted by the DPN network are directly collected using the capsule network,and a vector capsule of a fixed length is constructed by a convolution operation at the primary capsule layer.Finally,in the output capsule layer,a dynamic routing algorithm is used to gradually route the feature capsules to benign and malignant feature classes to obtain a fixed-length classification capsule.The output vector of the capsule contains different feature information of the nodules and the nodule length of the classification capsule.Probability representing benign and malignant classifications.In this paper,the LUNA16 data set is used as the main method,and the LIDC-IDRI data set is used as an auxiliary method to extract training data and doctor labeling.The data set is fully preprocessed to reduce redundant interference and improve training efficiency.Combining the advantages of Dual Path Network with fast feature extraction and high accuracy,and the capsule network’s sensitivity to spatial attitude and orientation information,a lung nodule classification prediction model is constructed which can obtain stable results with less training data and times.The experimental results show that in the LUNA16 data set,the classification prediction effect for benign and malignant nodules reaches 91.56%,which is close to the level of professional doctors.更多还原