基于多模态MRI图像的脑肿瘤分割方法

发布时间：2018-04-24 14:04

本文选题：脑肿瘤分割 + 多模态MRI图像　；参考：《电子科技大学》2017年硕士论文

【摘要】：脑肿瘤分割是一种从灰质、白质和脑脊液等正常的脑组织中,分离水肿、活跃和肿瘤坏死组织等不同的脑肿瘤结构的一种脑肿瘤辅助诊断技术。由于肿瘤在颅内产生,只能通过非侵入式显影方式(如:CT和MRI等)观察肿瘤的情况。在这些影像中肿瘤结构仅以灰度变化表征,不同设备不同条件获得的显影图像也可能有所差别,使得传统的图像分割方法很难较好地处理脑肿瘤图像的分割问题。随着MRI成像技术的不断成熟、多模态MRI图像的逐渐普及以及机器学习技术的飞速发展,脑肿瘤分割问题开始焕发新的活力。本文在这样的背景下,以脑肿瘤分割为立足点,以多模态MRI图像为基础,以机器学习方法为工具,针对脑肿瘤分割问题展开了如下研究:1、从脑肿瘤分割技术的发展脉络出发,在回顾经典方法的基础上,针对MRI图像本身的灰度不均匀的问题,利用多模态MRI图像数据本身的显影特异性,提出了一种针对多模态脑肿瘤分割的差分操作运算。同时,在多模态图像差分信息的辅助下,实现了利用多模态信息定位肿瘤区域的方法,为获得脑肿瘤的准确分割提供了一定的保障。从而也为充分的利用多模态MRI图像数据提供了一种可能的参考。在此基础上,将基于自组织神经网络(SOM)和主动轮廓模型(ACM)方法的自组织映射主动轮廓模型(SOAC)引入到脑肿瘤分割问题中,构造了基于多模态MRI图像的混合自组织主动轮廓模型(MSOAC)的脑肿瘤分割方法。以MRI数据预处理技术为先导,结合多模态差分定位肿瘤区域,通过MRI的slices数据为基础构建了一种基于多模态MRI图像的全自动的脑肿瘤分割方法。2、结合卷积神经网络和反卷积神经网络技术,以多模态MRI图像为依托,利用BRATS2015 training中3D MRI的slices数据,通过对要描述的肿瘤区域、肿瘤的活跃区域和肿瘤核区域分别构建反卷积子网络的方法,进行end-to-end的训练和参数调优,从而构建起多路深度反卷积网络模型。不同于以往在脑肿瘤分割中CNN采用的图像块训练的方法,这里直接利用slices训练,最后通过softmax预测每个像素的类别,得到pixel-wise的分割结果。
[Abstract]:Brain tumor segmentation is a technique for the diagnosis of brain tumors, which separates different brain tumor structures, such as edema, active and necrotic tissues, from normal brain tissues such as gray matter, white matter and cerebrospinal fluid. Because tumors occur in the brain, they can only be observed by non-invasive imaging (such as: Ct and MRI). In these images, the tumor structure is only represented by gray changes, and the development images obtained by different equipment and different conditions may also be different, which makes it difficult for traditional image segmentation methods to deal with the problem of brain tumor image segmentation. With the maturation of MRI imaging technology, the popularization of multimodal MRI images and the rapid development of machine learning technology, the segmentation of brain tumors begins to take on new vitality. In this context, this paper takes brain tumor segmentation as a foothold, based on multimodal MRI images, and uses machine learning method as a tool to study the problem of brain tumor segmentation as follows: 1, starting from the development of brain tumor segmentation technology. On the basis of reviewing the classical methods and aiming at the uneven grayscale of MRI images, a differential operation for multimodal brain tumor segmentation is proposed by using the development specificity of the multimodal MRI image data itself. At the same time, with the help of multi-modal image differential information, the method of using multi-mode information to locate tumor regions is realized, which provides a certain guarantee for accurate segmentation of brain tumors. It also provides a possible reference for making full use of multimodal MRI image data. On this basis, the self-organizing mapping active contour model (SOAC) based on the self-organizing neural network (SOM) and the active contour model (ACM) is introduced into the brain tumor segmentation problem. A new method of brain tumor segmentation based on multimodal MRI image is proposed, which is based on mixed self-organizing active contour model (MSOAC). Using MRI data preprocessing technology as the guide, combined with multi-modal differential localization of tumor region, Based on the slices data of MRI, an automatic segmentation method of brain tumor based on multimodal MRI image. 2. Combined with convolution neural network and deconvolution neural network technology, based on multimodal MRI image. Using the slices data of 3D MRI in BRATS2015 training, the deconvolution sub-network of the tumor region, the active region and the nuclear region of the tumor were constructed respectively to train and optimize the parameters of end-to-end. Thus the multichannel depth deconvolution network model is constructed. Different from the previous image block training method used by CNN in brain tumor segmentation, slices training is directly used here. Finally, the classification of each pixel is predicted by softmax, and the segmentation result of pixel-wise is obtained.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R739.41;R445.2;TP391.41

【相似文献】