面向认知演化的少儿DTI图像处理和统计分析方法研究

It is one of the most important issues in cognitive science to study the connection between structural development and cognitive evolution of human brain. We address this issue by investigating pediatric data of diffusion tensor imaging (DTI) across ages from 1 through 16 years old. We systematically study the problems of image processing and statistical analysis aiming at building the aforementioned connection. These problems include image processing techniques, e.g., tact-based registration, tractography and Establishment of white matter atlas, as well as those for statistical analysis, e.g., group analysis, correlation analysis and causal discovery. It is expected in this project that we propose image registration algorithms capable of normalizing DTI images with significant structural differences. We are also expecting to develop efficient probabilistic tractography algorithms that are able to deal with fiber crossing, and to generate various white mater atlases using sparse labels from experienced physicists. In addition, we attempt to establish computational models for cognitive evolution directly (instead of inspired) from structural development discovered by the statistical analysis in this project. We bridge studies on cognition and computation into a unified framework where we stress on computational issues arising from image processing and statistical analysis. Studies in this project do not only provide effective and powerful tool for fundamental research on cognition, but also play an important role in non-invasive diagnosis for cognitive deficiency as well as computer vision applications in practice.

本项目针对结构发育与认知演化的关联这一认知学的热点问题，以跨年龄段少儿（1-16岁）磁共振弥散张量成像(DTI)数据为研究对象,系统研究由结构发育影像建立认知演化模型的过程中涉及的图像处理和统计分析难点问题。项目研究基于神经纤维束的图像配准、纤维跟踪、神经纤维束模板构建，以及对比组分析、相关性和因果性分析等问题。预期提出将具有显著结构差异的DTI影像归一化的图像配准算法，合理处理纤维交叉的高效纤维跟踪概率算法，以及结合少量专家标记建立分年龄段纤维束模板的方法。此外，试图运用所研究的统计分析方法发现少儿大脑结构发育变化，并由此直接建立（而非启发）认知演化计算模型。项目以图像处理和统计分析等计算问题的研究为中心，同时在项目研究框架中有机融合了认知科学和计算科学。项目的研究将为发掘结构发育与认知演化的内在联系的基础研究提供技术支撑，同时对疾病无创性诊断和计算机智能处理等应用研究也有重要意义。

本项目针对结构发育与认知演化的关联这一认知学的热点问题，以跨年龄段少儿（1-16 岁）磁共振弥散张量成像(DTI)数据为研究对象,系统研究由结构发育影像建立认知演化模型的过程中涉及的图像处理和统计分析难点问题。通过项目研究，提出了能够克服幼儿神经束边界不明显、噪声较多等问题的纤维束聚类算法，能够自动高效地分割出主要神经束的方法；初步构建大脑主要的18条纤维束的概率模板；提出了能够结合图像表观和形状特征的基于回归的一系列图像对齐新方法；提出针对混合数据的高效在线学习方法。最后，基于提出的图像处理新方法，获得关于幼儿神经束发育的新的统计特性。共发表SCI论文21篇，EI 17篇。特别是研究成果获得国际多媒体旗舰会议ICME2015最佳学生论文。