特征与分数阶微分驱动的光流计算及其在大形变脑图像配准中的研究

Brain image registration with large deformation has been a technical difficulty in brain image registration due to its high complexity, low accuracy and low reliability. For brain images having the characteristics of complex structure and rich texture, this project will study on: (1) the mapping relation of feature point matching and image details preserving, which is taken as a constraint combined into optical flow model to realize the brain image registration with feature constraints; (2) fractional order differential computation applied to brain image, which is embedded into optical flow to enhance image gradient and improve registration precision of brain image with weak texture. The project will build feature and fractional order differential driven optical flow to realize high precision registration of brain images with large deformation. The research contents include: Effective feature extraction and description of large deformable brain image; High accurate matching algorithm based on improved K-d tree combined with random sampling uniformity; High precision registration model of feature driven optical flow for local details preserving; Fractional order calculus calculation on image processing; Feature descriptor algorithm based on fractional order differential; Optical flow model design based on fractional order differential calculation. This study not only has a great breakthrough in optical flow theory but also provides theoretical foundation and algorithm design for brain image registration with typical large deformation, which has academic significance and practical application value.

大形变脑图像配准因具有高复杂性、低精确度、低可靠性而成为脑图像配准技术中的难点。本项目针对脑图像的结构复杂、弱纹理丰富等特点，一方面研究特征点匹配与图像细节保持的映射模型，将其作为约束条件引入光流配准模型，实现特征约束的脑图像配准；另一方面研究脑图像分数阶微分模型，将其嵌入光流配准模型中进行梯度学习，提高弱纹理的脑图像配准精度；构建特征和分数阶微分共同驱动的光流配准模型，实现大形变脑图像的高精度配准。研究内容包括：大形变脑图像的有效特征提取与描述；基于改进的K-d树结合随机抽样一致性的高准确匹配算法；特征驱动下局部细节保持的高精度光流配准算法与分析；分数阶微积分的图像计算；基于分数阶微分的特征描述子算法；基于分数阶微分的光流模型计算。本研究不仅能在光流配准理论上有较大突破，而且对典型大形变脑图像配准提供理论基础和算法设计，具有学术意义和应用价值。

本项目是基于特征与分数阶微分约束的光流计算及脑图像的配准研究。通过对大形变脑图像边缘、拓扑等信息的准确计算，研究特征点匹配与图像细节保持的映射关系，将其作为特征约束条件引入光流配准模型，实现基于特征约束的脑图像配准；研究分数阶微积分的图像计算，嵌入光流配准模型中梯度学习，提高弱纹理的脑图像配准精度；构建特征与分数阶微分共同驱动的光流配准模型，实现大形变脑图像的高精度配准。研究小样本条件下的迁移学习的医学影像网络分析建模。经过四年的研究，本项目完成了特征与分数阶微分驱动的光流计算及其在大形变脑图像配准建模；构建了特征与分数阶微分驱动匹配模型，并将特征描述匹配融合配准模型的框架中，实现复杂条件下的高精度脑图像配准。.本项目在IEEE Transactions on Image Processing、IEEE Transactions on Circuits and Systems for Video Technology、IEEE Transaction on Information Forensics and Security、MICCAI、AAAI、ICASSP、ICIP、中国图象图形学报等国内外著名学术期刊和会议上发表受本项目资助论文16篇，其中SCI一区论文4篇，CCF-A/B会议论文5篇，SCI\EI检索16篇。申请发明专利8项，1项软件著作权，培养博士研究生2名，硕士研究生6名。项目负责人获得2021年度上海市优秀技术带头人，2019年上海市科技进步二等奖（排名第1）。.本项目深入研究了特征约束的光流理论，构建的特征与分数阶微分约束的配准模型用于脑图像结构配准，为高精度、快速及鲁棒性图像配准提供理论基础和关键技术。