【作者】 徐军；

【导师】 宋立新；

【作者基本信息】 哈尔滨理工大学 ， 信号与信息处理， 2020， 硕士

【摘要】 面对呈爆炸式增长的肺结节计算机断层扫描(Computed Tomography,CT)图像,能够迅速准确的对肺结节CT图像做出诊断,变成了一份艰难的工作。利用已经确诊的肺癌数据库中病例数据,通过肺结节CT图像检索,为医生提供了一种参考依据,从而降低误诊误判的几率。因此,肺结节CT图像的检索,在医生诊断过程中有着十分重要的意义。虽然传统的检索方法已经显著成效,但是仍有很多问题需要解决,而针对这些问题,本文研究内容如下:(1)针对原始数据对肺结节CT图像检索存在较大误差的问题,本文首先对肺结节CT图像进行预处理:主要包括病灶提取过程。通过实验对比最终采用最大类间方差法,形态学膨胀腐蚀,以及最大连通区域算法等方法实现肺结节CT图像的预处理过程,实验效果显著。(2)针对肺结节CT图像特征深层数据挖掘的问题,本文采用深度学习的方法对肺结节CT图像进行检索:主要采用基于深度置信网络(Deep Belief Network,DBNs)和迭代量化(Iterative Quantization,ITQ)搭建肺结节CT图像无监督哈希检索模型。实验结果表明,本方法在提高查准率和查全率等指标上效果明显。(3)针对肺结节CT图像数据库数据量不足的问题,本文采用迁移学习的方法对肺结节CT图像进行检索,主要采用监督学习哈希检索方法:基于Alex Net卷积神经网络(Convolutional Neural Networks,CNNs)模型,搭建迁移学习的网络模型,构造类似于局部敏感哈希(Locality-Sensitive Hashing,LSH)的哈希函数对图像进行二进制编码,然后采用分层检索的相似性度量方式,对肺结节CT图像进行检索。最终能够实现肺结节CT图像的监督学习哈希检索和分类并通过多个对比试验表明了本方法的优越性。最后针对LIDC-IDRI肺结节数据集对本文方法的检索性能进行评估,实验结果表明,本文所提出的两种基于深度学习的图像检索方法具有较优的图像特征提取和检索的能力,进一步验证了其在肺结节CT图像检索的有效性。更多还原

【Abstract】 Facing the explosive growth of computed tomography(CT)images of lung nodules,it is possible to quickly and accurately diagnose CT images of lung nodules,which has become a very difficult job.The method of retrieving CT images of lung nodules by computer simulation has been widely used in the diagnosis of lung cancer.Using the case data in the lung cancer database that has been diagnosed,through CT image retrieval of lung nodules,it provides a reference for doctors and reduces the chance of misdiagnosis and misjudgment.Therefore,the retrieval of CT images of pulmonary nodules is of great significance in the diagnosis of doctors.Although the traditional retrieval methods have achieved remarkable results,there are still many problems that need to be solved.In view of these problems,the research content of this article is as follows:(1)Aiming at the problem of large errors in CT image retrieval of lung nodules,this article first preprocesses the CT images of lung nodules: mainly including the process of extracting lesions,that is,extracting lung nodules from the CT images of lung.Through experimental comparison,the maximum inter-class variance method,morphological expansion and erosion,and the maximum connected area algorithm method are used to realize the preprocessing of CT images of pulmonary nodules.The experimental results are significant.(2)To solve the problem of deep data mining of CT image features of lung nodules,this paper uses deep learning to retrieve CT images of lung nodules: mainly based on Deep Belief Network(DBNs)and Iterative quantization,ITQ)to build an unsupervised hash retrieval model of CT images of lung nodules.The experimental results show that this method has obvious effects on improving the precision and recall rates.(3)Aiming at the insufficient amount of data in the CT image database of lungnodules,this paper uses transfer learning to retrieve lung nodule CT images,mainly using supervised learning hash retrieval methods: based on Alex Net convolutional neural network(CNNs)model To build a network model for transfer learning,construct a hash function similar to the Locally Sensitive Hash(LSH)to encode the image binary,and then use the similarity measure of hierarchical retrieval to retrieve CT images of lung nodules.Finally,the supervised learning hash retrieval and classification of CT images of lung nodules can be realized,and the superiority of this method is demonstrated through multiple comparative experiments.Finally,based on the LIDC-IDRI lung nodule data set,the retrieval performance of this method is evaluated.The experimental results show that the two deep learning-based image retrieval methods proposed in this paper have better image feature extraction and retrieval capabilities,and further verify The effectiveness of CT image retrieval in pulmonary nodules was demonstrated.更多还原