【作者】 李琪；

【导师】 贾同；

【作者基本信息】 东北大学 ， 控制工程（专业学位）， 2019， 硕士

【摘要】 肺癌是我国肿瘤疾病中的头号杀手,据统计5年生存率为15%。早发现、早诊断、早治疗可以大大提高肺癌患者的生存率。肺结节图像是肺癌的早期重要临床特征,具有体积小、分布不均匀、形状不规则的特点。随着CT技术的快速发展,在提高CT图像清晰度的同时数据量也呈爆炸性增长,通常每位病人每次扫描就有成百甚至上千张CT图像,有经验的医师也难以做到快速准确的判断。而且检测结果经常会受主治医师主观因素影响,出现误诊、漏诊现象在所难免。在这一背景下,基于CT图像的肺结节计算机辅助检测技术已经成为研究热点,可将CT图像中所包含的结节区域自动标注出来,提醒医师重点查看,从而减少医生工作量,有效提高结节检测的准确率和效率。本文基于肺部CT图像,依据计算机辅助检测诊断的思路,提出了基于聚合Unet深度网络的肺结节检测技术。该技术主要分为三部分,分别为肺实质分割、肺结节的分割和肺结节的分类。首先,基于Kmeans聚类算法和形态学操作提取肺实质区域,缩小结节检测区域,提高检测速度;然后,提出一种基于聚合Unet深度网络的肺结节分割算法,建立疑似结节集;在此基础上,提出一种基于迁移学习和支持向量机的肺结节分类算法,进一步提高检测的准确率。本文的主要研究内容如下:一、基于Kmeans的肺实质分割。基于传统阈值算法进行肺实质分割时,由于数据集中CT采集设备不同,造成同病异影现象,难以找到最优阈值,因此本文采用Kmeans算法完成准确的肺实质分割,同时采用形态学操作算法对肺实质进行修补,使其分割更加完整,实验结果表明Kmeans聚类和形态学操作算法可以完成肺实质的准确分割。二、基于聚合Unet网络的肺结节分割。本文针对结节图像的特点,提出一种聚合Unet网络,其可以将浅层的边缘、颜色、纹理等信息和深层的抽象特征更好的融合。本文采用Luna16肺部CT数据集进行验证,实验结果表明聚合Unet网络针对肺结节分割,其分割准确率和训练时间优于Unet与Segnet网络。三、基于迁移学习的肺结节分类。针对医学图像数据集小,很难在有限的数据集上训练出能准确提取结节特征的有效深度网络。本文基于迁移学习的方法,将在大量的自然图像数据集下训练得到的卷积神经网络模型参数进行调整,并基于支持向量机方法对提取的特征进行真假阳性分类。实验结果表明,基于迁移学习和支持向量机的方法可以准确的分类肺结节。更多还原

【Abstract】 Lung cancer is the number one killer of cancer diseases in China.According to statistics,the 5-year survival rate is 15%.Early detection,early diagnosis and early treatment can greatly improve the survival rate of lung cancer patients.The pulmonary nodule image is an important early clinical feature of lung cancer,and it has the characteristics of small volume,uneven distribution and irregular shape.With the rapid development of CT technology and image resolution improvement,CT has also brought a lot of data,per patient per scan usually have hundreds or even thousands of sheets CT image.It’s also difficult for experimental doctors to achieve fast and accurate judgment.In this context,computer-aided detection of lung nodules based on CT images has become a research hotspot.The nodule regions contained in CT images can be automatically labeled to remind doctors to focus on viewing,thus reducing the workload of doctors and effectively improving the accuracy and efficiency of nodule detection.Based on the CT image of the lung and the idea of computer-aided detection and diagnosis,this paper proposes a lung nodule detection technique based on the aggregated Unet deep network.The technique is mainly divided into three parts,namely,the segmentation of lung parenchyma,the segmentation of pulmonary nodules,and classification of pulmonary nodules.First,we extract the lung parenchyma area based on Kmeans clustering algorithm and morphological operation,this part can reduce the nodule detection area,and improve the detection speed;On this basis,to further improve the accuracy of detection,we propose a lung nodule classification algorithm based on migration learning and support vector machine.The main research contents of this paper are as follows:First,Segmentation of lung parenchyma based on Kmeans method.When the lung parenchymal segmentation is based on the traditional threshold algorithm,it is difficult to find the optimal threshold because of the different CT acquisition devices in the dataset.Therefore,the Kmeans algorithm is used to complete the accurate segmentation of lung parenchyma,and the morphological operation algorithm is used.The lung parenchyma was repaired to make the segmentation more complete.The experimental results show that Kmeans clustering and morphological manipulation algorithms can complete the accurate segmentation of lung parenchyma.Second,Segmentation of pulmonary nodules based on aggregated Unet networks.In this paper,based on the characteristics of nodule images,a kind of aggregated Unet network is proposed,which can better integrate the shallow edge,color,texture and other deep abstract features.This paper uses 888 Lunal 6 lung CT dataset to verify.The experimental results show that the aggregated Unet network for lung nodules segment,its segmentation accuracy and training time are better than Unet and Segnet structures.Third,Classification of lung nodules based on transfer learning.For medical image datasets,it is difficult to train an effective depth network that can accurately extract nodule features on a limited data set.Based on the transfer learning method,this paper fine-tunes the convolutional neural network model parameters trained under a large number of natural image data sets,and uses the support vector machine to complete the two classifications to reduce false positives.The experimental results show that the algorithm based on transfer learning and support vector machine can accurately classify lung nodules.更多还原