基于机器学习的人脑MR图像结构区域层次化自动标记方法研究

项目名称： 基于机器学习的人脑MR图像结构区域层次化自动标记方法研究

项目编号： No.81471733

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 医药、卫生

项目作者： 王乾

作者单位： 上海交通大学

项目金额： 70万元

中文摘要： 脑部结构区域自动标记是人脑MR图像分析中亟需解决的问题，其性能对后续应用、分析有决定性的影响。本项目利用机器学习的方法实现在MR图像中对多个人脑结构区域的自动标记。我们综合考虑区域标记的三个主要难点，即：（1）如何为待标记图像选择最佳分类器（图谱）；（2）如何设计分类器以将标记信息由图谱传递至待标记图像；（3）如何将来自多个分类器（图谱）的标记结果加以动态融合。我们提出一种创新的层次化结构区域标记系统，将上述三个问题的解决置于统一的框架之中。在处理新的待标记图像时，我们首先根据该图像呈现的内容选择最佳分类器；在将选中的分类器作用该图像后，我们对多分类器的输出进行融合。在较低层分类器上获得的标记将被作为较高层的输入；这样，层次化系统可以逐步优化待标记图像中的结构区域分布。我们的方案建立在前期工作基础之上，技术上可行，性能上先进，有望成为人脑MR图像结构区域标记的主要手段之一。

中文关键词： 医学图像分割；医学图像处理；磁共振成像；结构区域标记；机器学习

英文摘要： Automatic labeling of structural regions in human brain MR images is fundamentally necessary to image analysis. The performances of region labeling are crucial to the subsequent applications and statistical analyses. In this project, we aim to propose an automatic labeling method of multiple structural regions in human brains by utilizing the machine learning technique. We systematically consider all three major obstacles in terms of region labeling, i.e., (1) how to select the best classifiers/atlases for the to-be-labeled image; (2) how to design and apply the classifiers such that the labeling information could be propagated from atlases to the to-be-labeled image; (3) how to determine the optimal fusion of outputs from multiple classifiers/atlases. To this end, we propose a novel hierarchical labeling system, and incorporate all three points above into a unified framework. When a new to-be-labeled image comes, we first select the best classifiers based on the content rendered by the new image. After applying only selected classifiers, we are capable of fusing outputs from multiple classifiers dynamically. The result generated by a lower level in the hierarchical system is regarded as the input to an upper level, where the labeling result is further refined. Our proposal is fully based on the substantial works which have been conducted in our researches already. We believe that our technique is innovative and owns promising prospect when applied to the automatic label of brain regions in MR images.

英文关键词： Medical Image Segmentation;Medical Image Processing;Magnetic Resonance Imaging;Structral Region Labeling;Machine Learning