基于深度学习的乳腺癌早期诊断研究

发布时间：2018-01-27 23:32

本文关键词： 深度学习乳腺癌动态增强核磁共振成像三维卷积神经网络迁移学习　出处：《杭州电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：在世界范围内,乳腺癌以其高发病率、高死亡率严重威胁女性身体健康,且近年来乳腺癌新增病例数持续上升。因其发病机理不确定和病情隐匿,使得早期乳腺癌很难被发现。乳腺动态增强核磁共振影像(Dynamic Contrast-enhanced magnetic resonance imaging,DCE-MRI)具有较高的软组织分辨能力,近年来成为广泛应用于乳腺癌早期检查的重要方式。模仿生物视觉原理的深度学习能够自动学习数据层次化的特征,使其在图像、语音、自然语言处理等方面取得巨大成就。人们不断探索将深度学习应用在生物医学领域,并取得了一定成果。本论文的研究工作主要是基于深度学习方法在早期乳腺癌诊断上的应用,探索研究基于不同深度学习方法,并利用不同模态的影像数据对乳腺癌早期诊断。本文的主要研究内容如下:(1)基于非监督学习堆叠自编码特征提取方法与分类利用非监督堆叠自编码(Stacked Autoencoder,SAE)进行乳腺癌特征提取与分类。首先对实验数据预处理,提取ROI及PCA白化,其次利用非监督逐层训练的方式提取不同层级的特征,最后利用Softmax进行早期乳腺癌良恶性判别。(2)基于三维卷积神经网络乳腺癌早期诊断提出构建三维卷积神经网络,通过平移、旋转、镜像等方式将样本数据扩充。然后利用二维和三维卷积网络来对疾病分类预测。利用三维增强影像序列和增强率影像实现早期乳腺癌的识别分类。(3)迁移学习判别乳腺癌良恶性研究研究迁移学习模型在早期乳腺癌良恶性分类上的应用,将在大数据集(如ImageNet)上预训练得到的网络模型作为底层和中层的特征提取器,迁移到MRI影像数据集,微调模型参数进行分类。从研究结果来看,本文所提出的研究方法能够较好的对早期乳腺癌识别分类。降噪堆叠自编码对乳腺癌早期诊断实验结果AUC达到0.85。基于3DCNN早期乳腺癌分类实验结果AUC值0.80,灵敏度和特异性分别达到0.82和0.74。在迁移网络模型分类实验中,AUC,灵敏度和特异性分别达到0.86,0.85和0.81。未来若将深度学习方法自动学习到的不同层级的特征融合到MRI辅助诊断系统中,将有助提升系统的性能,具有较好应用前景。
[Abstract]:In the world, breast cancer is a serious threat to the health of women because of its high incidence and high mortality, and in recent years, the number of new cases of breast cancer has been increasing, because of its pathogenesis of uncertainty and hidden disease. This makes early breast cancer difficult to detect. Dynamic enhanced Magnetic Resonance Imaging (. Dynamic Contrast-enhanced magnetic resonance imaging. DCE-MRI) has high soft tissue resolution. In recent years, it has become an important way to be widely used in early detection of breast cancer. Deep learning imitating biological vision principles can automatically learn the hierarchical features of data and make it in images and speech. Great achievements have been made in natural language processing and so on. People are constantly exploring the application of in-depth learning in biomedical fields. The research work in this paper is mainly based on the application of depth learning method in the diagnosis of early breast cancer, and the research is based on different depth learning methods. The main contents of this paper are as follows: (1) based on unsupervised learning stack self-coding feature extraction method and classification using unsupervised stack self-coding (. Stacked Autoencoder. First, ROI and PCA whitening were extracted from the experimental data, and then the features of different levels were extracted by unsupervised training. Finally, Softmax is used to distinguish benign and malignant breast cancer. (2) based on three-dimensional convolution neural network for early diagnosis of breast cancer, a three-dimensional convolutional neural network is proposed to construct a three-dimensional convolutional neural network, which can be translated and rotated. Image image is used to expand the sample data, and then two dimensional and three dimensional convolution networks are used to predict disease classification. Three-dimensional enhanced image sequences and enhancement rate images are used to realize the recognition and classification of early breast cancer. Study on differentiation of benign and malignant Breast Cancer by Migration Learning; Application of Migration Learning Model in Classification of benign and malignant Breast Cancer in early stage. The network model pre-trained on the big data set (such as Image net) is used as a feature extractor at the bottom and middle levels to migrate to the MRI image data set. Fine tuning the model parameters to classify. From the results of the study. The research method proposed in this paper can identify and classify early breast cancer better. The experimental results of early breast cancer diagnosis based on noise reduction stacking self-coding can reach 0.85. The classification of early breast cancer based on 3DCNs is true. The AUC value was 0.80. The sensitivity and specificity were 0.82 and 0.74.The sensitivity and specificity of AUC were 0.86 respectively. 0.85 and 0.81. In the future, if the features of different levels of the automatic learning method are fused into the MRI aided diagnosis system, it will help to improve the performance of the system and have a good prospect of application.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;R737.9

【参考文献】