基于卷积神经网络的肺结节自动检测深度学习模型

发布时间：2018-03-10 14:36

本文选题：肺结节　切入点：CT图像　出处：《太原理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：近日,雾霾再次出现并严重影响着人们的生活,同时危害着人们的身体健康。在雾霾、PM2.5等大气污染的影响下,全国乃至全球范围内肺癌患者的数量呈现出指数增长趋势。肺癌中病灶体称之为肺结节,它具有体积较小、形态各异、多与胸腔内膜粘连等特点,同时又有支气管、血管等干扰,导致对肺癌的早期诊断具有一定的难度。同时,肺结节在肺部所占面积较小,医师仅用肉眼对CT图像观察,根据已有知识和自身经验找出病灶体肺结节并对其进行良恶性判断容易造成误诊或者漏诊。在低剂量CT薄层扫描的技术广泛应用后,影像数据爆炸式增长与人工诊断力量严重不足的矛盾,大数据和数据分析技术发展的不协调,都有可能导致肺癌诊断准确率的降低。随着计算机技术的发展和应用,在众多大型医院中,医师都借助于计算机对肺癌进行辅助诊断。在较为完善的计算机辅助诊断CAD中,通过图像预处理、分割、特征提取和特征选择等步骤,最终实现对肺结节的分类。提高肺结节良恶性分类的准确率是最终的目的,而特征提取是关键的步骤。本文在对国内外研究现状进行研究后,提出了存在的问题还解决办法。针对大数据的CT图像,建立基于卷积神经网络的肺结节良恶性自动诊断模型,主要的研究工作包括以下几个方面:1.针对计算机辅助诊断系统中复杂的算法和人为干扰产生的不可抗的因素,本文在利用区域生长对CT图像进行简单预处理后得到肺实质图像,通过双线性差值存储为样本。利用样本对自定义的卷积神经网络模型进行训练并达到对肺结节良恶性诊断的目的。此方法可以在避免特征提取等复杂算法的基础上,提高肺结节分类准确性和分类速度。2.特征作为肺结节主要的分类前提,提取特征是必不可少的步骤。在传统方法中,特征提取方法都是根据经验进行人为设定,包括灰度、形状以及纹理等底层特征,但是这些特定的特征有一定的局限性。本文直接将图像作为原始样本输入到卷积神经网络内,通过隐含层的自主学习提取到关键特征。已有的方法中仅将最后一层输出作为特征,忽略了隐含层的特征。由于每个特征的贡献率不同,且经过多层学习可能在最后一层丢失,因此本文将每层特征经过PCA降维得到最终的融合特征。虽然融合特征无法有确切的描述,但是通过分类器可以得到较为准确的分类结果。本文通过搭建卷积神经网络模型,在大数据的样本的试验认证下,在降低了算法的复杂性的同时提高了整体肺结节的检出率,降低了误诊率和漏诊率。这为医师的诊断提供了更为准确、有效和方便的方法,对肺癌的早期诊断和治疗有着积极作用。
[Abstract]:Recently, haze has reappeared and seriously affected people's lives and their health. Under the influence of air pollution such as haze and PM2.5, The number of lung cancer patients in the whole country and the whole world shows an exponential increasing trend. The focus body of lung cancer is called pulmonary nodule. It has the characteristics of small size, different shape, more adhesion with thoracic endomembranium, and also has bronchus. The early diagnosis of lung cancer is difficult due to the interference of blood vessels. At the same time, the area of pulmonary nodules in the lung is small, so doctors only use the naked eye to observe CT images. It is easy to misdiagnose or miss diagnosis of pulmonary nodules based on their own knowledge and experience. The contradiction between the explosive growth of image data and the serious shortage of artificial diagnostic power, the incoordination between big data and the development of data analysis technology, may lead to a decrease in the diagnostic accuracy of lung cancer. With the development and application of computer technology, In many large hospitals, doctors help diagnose lung cancer with the aid of computer. In CAD, image preprocessing, segmentation, feature extraction and feature selection are used. Finally, the classification of pulmonary nodules is realized. Improving the accuracy of benign and malignant classification of pulmonary nodules is the ultimate goal, and feature extraction is the key step. According to big data's CT image, the automatic diagnosis model of benign and malignant pulmonary nodules based on convolution neural network is established. The main research work includes the following aspects: 1. In view of the complex algorithm and the indelible factors caused by human interference in the computer-aided diagnosis system, the lung parenchyma image is obtained by using the region growth to preprocess the CT image. The bilinear difference is stored as the sample. The self-defined convolution neural network model is trained by the sample and the diagnosis of benign and malignant pulmonary nodules is achieved. This method can avoid complex algorithms such as feature extraction and so on. Improve the accuracy and speed of classification of pulmonary nodules. 2. Feature is the main premise of classification of pulmonary nodules, feature extraction is an essential step. In traditional methods, feature extraction methods are artificially set according to experience, including gray level. In this paper, the image is directly input into the convolutional neural network as the original sample. The key features are extracted by autonomous learning of the hidden layer. In the existing methods, only the output of the last layer is taken as the feature, while the feature of the hidden layer is ignored. Because the contribution rate of each feature is different, and the multi-layer learning may be lost in the last layer, Therefore, the final fusion feature can be obtained by reducing the dimension of each layer by PCA. Although the fusion feature can not be described exactly, the classification result can be obtained by classifier. In this paper, a convolution neural network model is built. Under the experimental verification of big data's sample, the complexity of the algorithm is reduced and the detection rate of global pulmonary nodules is increased, and the misdiagnosis rate and missed diagnosis rate are reduced. This provides a more accurate, effective and convenient method for the diagnosis of doctors. It plays an active role in the early diagnosis and treatment of lung cancer.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R734.2;TP391.41

【参考文献】