基于深度卷积神经网络的医学图像肺结节检测方法研究

发布时间：2018-03-24 01:12

本文选题：肺结节检测　切入点：卷积神经网络　出处：《江南大学》2017年硕士论文

【摘要】：随着时代的发展,空气污染的加重,肺癌已经成为威胁人类生命恶性程度最高的肿瘤之一。及早发现和治疗可以大大地提高病人的存活率。肺癌的早期症状在医学影像上大多表现为孤立的肺结节,在胸片上通常表现为圆形或近似圆形的低对比度光斑,没有特殊的处理,肉眼是很难将肺结节与肺部其它的软组织区分出来。深度学习是机器学习的一个新兴领域,在近年来得到了高速的发展。深度学习其实就是通过组建多个隐层神经网络和使用大量的数据进行训练,从而提取出这些数据中更为有用的特征来提高模型的预测或分类的准确性。它在图像处理、语音和自然语言处理等多个领域都取得了不错的成果。本文主要研究针对X光胸片的肺部结节自动检测方法,并初步探索了应用深度学习方法在CT图像上进行肺结节检测。主要内容如下:(1)通过研究传统胸片肺结节检测方案,提出一种基于卷积神经网络的胸片肺结节检测方案。该方案首先对胸片进行预处理,用USM锐化的方法对图像中的结节信号进行增强。然后在胸片上用滑动窗口的方法切取小块,下采样后输入进预先训练好的卷积神经网络进行分类,得到整张胸片的候选结节区域。最后根据面积阈值排除掉大量的假阳性。在JSRT数据库上的实验结果表明,该方法在相同的假阳性水平下比相关文献中方法可以检测出更多的肺结节。(2)通过研究肺结节图像小块下采样后的形态表现,提出一种集成卷积神经网络用于胸片上的肺结节检测。该方法也是先对胸片进行预处理,用USM锐化方法对图像中的结节信号进行增强。然后用滑动窗口的方法从胸片上切取229×229的小块,接着分别下采样到12×12,32×32和60×60三个不同的尺度,分别输入进三个预先训练好的不同的卷积神经网络,最终的分类结果由这三个神经网络的输出结果投票决定。由此可以得到整个胸片的候选区域,最后根据面积阈值排除大量的假阳性。在JSRT数据库上的实验结果表明,该集成方法可以排除掉大量的假阳性,使其在相同的假阳性水平下比相关文献及上一章的方法检测出更多的肺结节。(3)初步探索了在CT图像上进行疑似肺结节的检测。从肺结节的立体特性出发,提出一种多输入的卷积神经网络模型用于CT图像上的疑似肺结节检测。该方法首先对CT切片进行预处理,用USM锐化增强结节的信号。然后在相邻的切片的相同位置切取相同大小的小块,输入进预先训练好的多输入卷积神经网络,得到整个CT序列中的候选区域。用面积阈值可以做初步的假阳性筛选工作。在从LIDC-IDRI中选取的子集数据库上的实验结果来看,结节的检出率满足要求,可以用于对其做更加深层次的研究。
[Abstract]:With the development of the times, the air pollution is getting worse. Lung cancer has become one of the most malignant tumors threatening human life. Early detection and treatment can greatly improve the survival rate of patients. On chest radiographs, they usually appear as round or nearly circular low-contrast spots, without special treatment, and it is difficult for the naked eye to distinguish pulmonary nodules from other soft tissues of the lungs. Deep learning is a new field of machine learning. In recent years, rapid development has been achieved. In fact, deep learning is training by building multiple hidden layer neural networks and using a large amount of data. To extract more useful features from these data to improve the accuracy of the prediction or classification of the model, which is used in image processing, Many fields, such as speech processing and natural language processing, have achieved good results. This paper mainly studies the automatic detection method of pulmonary nodules based on X-ray chest radiographs. The paper also preliminarily explored the application of deep learning method to the detection of pulmonary nodules on CT images. The main contents are as follows: 1) by studying the traditional chest radiographic lung nodule detection scheme, A new method for detecting pulmonary nodules in chest radiographs based on convolutional neural network is proposed. Firstly, the chest radiographs are preprocessed, and the nodule signals in the images are enhanced by USM sharpening, and then the small pieces are cut out by sliding window on the chest radiographs. After sampling, we input the pre-trained convolution neural network to classify the candidate nodule area of the whole chest film. Finally, a large number of false positives are excluded according to the area threshold. The experimental results on JSRT database show that, This method can detect more pulmonary nodules at the same false positive level than in related literature. An integrated convolution neural network is proposed to detect pulmonary nodules on chest radiographs. USM sharpening method was used to enhance the nodule signal in the image, and then the small pieces of 229 脳 229 were cut from chest radiographs by sliding window method, and then sampled to three different scales of 12 脳 12, 32 脳 32 and 60 脳 60, respectively. The final classification results are determined by the output results of the three neural networks, and the candidate regions of the whole chest radiography can be obtained by the input of three pre-trained different convolutional neural networks, and the final classification results are determined by the output results of the three neural networks. Finally, a large number of false positives are excluded according to the area threshold. The experimental results on the JSRT database show that the integration method can eliminate a large number of false positives. At the same false positive level, more pulmonary nodules were detected under the same false positive level than those in the previous chapter. (3) A preliminary exploration was made for the detection of suspected pulmonary nodules on CT images. A multi-input convolution neural network model is proposed for detecting suspected pulmonary nodules on CT images. USM sharpening is used to enhance the signal of nodules. Then the same size small pieces are cut at the same position of adjacent slices and input into the pre-trained multi-input convolution neural network. The candidate regions in the whole CT sequence are obtained. The initial false positive screening can be done by using the area threshold. The experimental results from the subset database selected from LIDC-IDRI show that the detection rate of nodules meets the requirements. It can be used for deeper research.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R734.2;TP391.41

【参考文献】