基于多模态医学影像的阿尔兹海默症（AD）的计算机辅助诊断是脑影像智能计算领域的研究热点和前沿。但该问题是一个典型高维度小样本的机器学习问题，且各模态的互补信息的有效利用没有得到很好的解决。本项目围绕上述问题展开深入研究，提出新型的多模态图像特征提取方法：（1）研究图像灰度标准化对AD诊断的影响，通过在图册模型中对待分割的体素进行加权约束，在完成分割的同时，充分利用图册的信息，完成待分割图像的灰度标准化；综合考虑灰度标准化对图像分割的影响，提出基于图册表观模型的联合脑结构分割与灰度标准化方法。（2）针对高维度小样本的问题，研究基于多线性模型的深度卷积网络分类方法：利用正则化的多线性回归模型获得图像的目标级表示，通过考虑目标图像每一个体素点对应的局部区域的相似度，对每一个体素点都赋予一个不同的权重；同时考虑每个块邻域之间的相似度，赋予不同的权重获得图像的目标级表示，作为深度网络的输入，完成图像降维。同时使用多线性模型综合考虑各个模态的互补信息，进而完成AD图像的分类。在此基础上，研究基于多示例学习的AD 诊断方法：利用T-test 方法，通过对像素点判别能力的排序，生成每种模态图像的标志点；以标志点为中心，取不同大小的块为基础，计算纹理等特征，计算标志点的多尺度特征表示；每个标志点的特征表示作为一个示例，完成图像的多示例表示。最后，探索三维的卷积神经网络方法，充分利用图像的结构信息，降低图像在从三维到二维模型中的结构信息的丢失。初步的实验结果表明，本项目提出方法的分类结果优于现有方法。本项目的研究内容对于AD临床诊断技术水平的提高、脑影像智能计算的发展都具有重要意义。

Computer-aided diagnosis of Alzheimer's disease (AD) via multi-Modality medical imaging is the hot research topic and frontiers of intelligent brain imaging analysis. However, it is a typical high dimensional and small sample machine learning problem, and the effective use of the complementary information on each modality is not solved well. This project makes a deep research on the above challenge and proposes a novel multi-modality image feature extraction method. First, we explore the effect of gray scale standardization on diagnosis of AD by restricting weight on the segmented voxels in the album model, and making full use of the information of the atlas while completing the segmentation and gray standardization of the image to be segmented. The method of joint brain structure segmentation and gray standardization is proposed based on the atlas model by the effect of gray standardization on image segmentation. Then, deep convolution network classification is studied based on the multi-linear model aiming at the problem of high-dimensional small samples. The target-level representation of the image is obtained by using a regularized multilinear regression model. We use the similarity of voxel on partial area to give the different weight on every voxel. At the same time, the similarity between the neighborhoods of each block is assigned different weights to obtain the target-level representation of the image as the input of the deep network and complete the image dimensionality reduction. The multi-linear model is used to comprehensively consider the complementary information of each modality, and then the classification of AD images is completed. Based on the above research, multi-instance learning is expended on AD diagnostic. T-test is used to sort the discrimination ability and generate the marker points of each modal image. Based on the marker points, different sizes blocks are used to calculate texture and multi-scale feature representation of landmarks. The feature representation of each marker point is seen as an example then we complete the multiple example representation of the image. To make full use of the image structural information and reduce the loss of structural information in images from 3D to 2D models, we try to explore the three-dimensional convolution network method. Preliminary experimental results show that the classification results of the proposed method are better than the existing methods. The project is of great significance for the improvement of the clinical diagnostic techniques of AD and the development of intelligent brain imaging computing.