卷积神经网络在乳腺肿块分类中的研究与应用

发布时间：2018-02-09 04:40

本文关键词： 乳腺癌乳腺X线图像计算机辅助诊断卷积神经网络迁移学习　出处：《昆明理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：乳腺癌是女性最常见的恶性肿瘤之一,女性一生中患乳腺癌的可能性约为10%,在所有女性恶性肿瘤中,乳腺癌的发病率已居于首位,同时其致死率高达40%以上。目前对乳腺癌尚无积极的预防手段,早期诊断和及时治疗是提高乳腺癌术后生存率的唯一途径。但是乳腺肿块的研究仍存在一定的难点,因为乳腺肿块大小和对比度差异性较大,而且易受伪影和周围腺体组织的干扰,极大地影响诊断系统的精度。本文在对目前乳腺癌计算机辅助诊断系统研究的基础上,深入研究了卷积神经网络模型在乳腺肿块分类中的应用问题,开展的主要研究工作如下:(1)乳腺肿块图像去噪。为了尽可能消除噪声对乳腺X线图像的影响,同时又保证乳腺肿块的边缘信息,本文分析了乳腺X线影像中噪声的特点,对比多种图像去噪算法,并在乳腺X线影像上进行试验,经实验结果确定使用小波算法去除图像噪声。(2)肿块区域分割与形态学处理。论文在乳腺肿块的分割方面,应用大津算法对乳腺X线影像进行分割,提取出感兴趣区域。然后运用形态学方法对粗分割得来的感兴趣区域进行膨胀与开闭运算处理,用于保存肿块图像的边缘信息和消除肿块内部的孔洞,得到最终的肿块图像。(3)肿块分类。论文选择使用迁移学习的方法,将大规模深度卷积神经网络应用在乳腺肿块良恶性区分中。论文对卷积神经网络模型结构和训练过程进行分析与研究,尝试使用迁移学习的方法,将在自然图像集上训练完毕的GoogLeNet和AlexNet在乳腺肿块图像上进行微调,微调后的模型应用在乳腺肿块的分类中,实现了基于卷积神经网络的乳腺肿块良恶性的区分。文中还对比从乳腺图像上直接训练的浅层卷积神经网络和人工设定特征实现分类的几种模型,实验结果也表明了基于迁移学习的深度卷积神经网络模型在乳腺肿块良恶性区分中存在较大的优势。
[Abstract]:Breast cancer is one of the most common malignant tumors in women. At the same time, the fatality rate of breast cancer is over 40%. At present, there is no positive preventive method for breast cancer. Early diagnosis and timely treatment are the only way to improve the survival rate of breast cancer. However, there are still some difficulties in the study of breast masses. Because the size and contrast of breast masses vary greatly, and they are easily disturbed by artifacts and surrounding glands, the accuracy of the diagnostic system is greatly affected. In this paper, the application of convolution neural network model in classification of breast masses is deeply studied. The main research work is as follows: 1) De-noising of breast masses. At the same time, the edge information of breast mass is guaranteed. This paper analyzes the characteristics of noise in mammography, compares various image denoising algorithms, and carries out experiments on mammography. The experimental results confirm that the wavelet algorithm is used to remove the noise of the image and to deal with the morphology of the mass. In the aspect of the segmentation of the breast mass, the paper applies the Otsu algorithm to segment the mammary mammary X-ray image. The region of interest is extracted, and then the rough segmentation of the region of interest is processed by the morphological method, which is used to preserve the edge information of the mass image and eliminate the holes inside the mass. The final mass image is obtained. The large scale deep convolution neural network is applied to distinguish benign and malignant breast masses. This paper analyzes and studies the model structure and training process of convolution neural network, and tries to use the transfer learning method. The GoogLeNet and AlexNet trained on the natural image set are fine-tuned on the breast mass image, and the fine-tuned model is applied to the classification of the breast mass. The classification of benign and malignant breast masses based on convolution neural network is realized. The experimental results also show that the deep convolution neural network model based on migration learning has a great advantage in differentiating benign and malignant breast masses.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R737.9;TP391.41

【参考文献】