基于血管关键点和旋转不变自相似特征的多模态眼底图像稳健配准研究

Red Free (RF) / Fluorescein Angiographic (FA) image registration is a vital step in multimodal fusion based precise diagnosis and laser photocoagulation for fundus. However, there are significant nonlinear intensity difference between these two types of images, thus the traditional keypoint detection method and feature descriptor are compromised, challenging the robustness of the registration. Based on the geometric consistency of the vessel in the multimodal images, a registration method using vascular keypoints and rotation-invariant self-similarity feature is proposed in this project. The stable geometric features, such as local bending, orientation change and bifurcation of the vessel are detected as keypoints to improve the repeatability and stability. Furthermore, the self-similarity idea is introduced into the local descriptor, and the spatial search region, self-similarity measure and normalized methods are investigated to extract feature vector based on the local intensity-pattern’s repetitions, instead of directly using the image intensity. The previously proposed main orientation list is adopted to achieve distinctiveness, rotational invariance along with robustness to non-linear intensity difference for the feature vector, so as to improve the overall robustness of the registration. In this project, for multimodal fundus image registration, a rotation invariant self-similarity feature is proposed for the first time to improve the performance of image registration, promoting the appearance of new ideas for the universal scientific problem of "modal independent invariant feature extraction" in multimodal image registration.

无赤光/造影图像配准是多模融合的眼底精确诊断、光凝手术导航中的关键一环，这两类图像间存在显著的非线性灰度差异，使常规标志点检测与描述方法的性能受到制约，给配准的稳健性带来挑战。本项目根植于图像间血管的几何结构一致性，提出一种基于血管关键点和旋转不变自相似特征的配准方法，将血管的局部弯曲、方向改变、分支等稳定存在的几何关键点作为标志点，提高检测的重复性与可靠性，同时将自相似思想引入到局部特征描述中，通过研究搜索邻域、自相似度量和归一化方法，建立基于局部灰度模式重复性而非直接基于灰度的特征描述，结合申请人前期提出的主方向链表，实现特异性、旋转不变性和对非线性灰度差异鲁棒性的结合，提高配准稳健性。本项目针对多模态眼底图像，首次提出旋转不变的自相似特征描述，在助力解决眼底图像配准具体问题的同时，也将为“模态无关的不变特征提取”这一多模态图像配准面临的共性科学问题开辟新思路。

盲和视力损伤严重影响人民群众的身体健康和生活质量，加重了家庭和社会负担，是重大的公共卫生问题。截至2013年，我国视力残疾患者增至1300万人，年均新增45万盲人、135万低视力患者，每分钟即新增1个盲人、3个低视力患者；从导致视力残疾的因素看，白内障、青光眼、老年黄斑变性、糖尿病视网膜病变是四大致盲诱因。无赤光/造影图像配准是多模融合的眼底精确诊断、光凝手术导航中的关键一环，这两类图像间存在显著的非线性灰度差异，使常规标志点检测与描述方法的性能受到制约，给配准的稳健性带来挑战。本项目根植于图像间血管的几何结构一致性，提出一种基于血管关键点和旋转不变自相似特征的配准方法，内容包括：（1）研究几何显著的血管关键点检测方法，提高检测的重复性与可靠性。（2）建立旋转不变的自相似特征描述，实现特异性、旋转不变性和对非线性灰度差异鲁棒性的结合。（3）研究利用极大似然估计采样一致性消除误配，完成可靠的变换模型估计。（4）依据国际公认的评价标准，进行配准结果评估。本项目从深圳市南山医院采集1200幅多模态眼底影像，采用均方误差(RMSE)对配准算法进行误差分析，实验结果表明，本项目提出的基于旋转不变距离的配准方法误差在1个像素左右，匹配成功率为百分之九十以上，性能超过国内外同类方法，且配准性能不受图像旋转角度的影响。本项目针对多模态眼底图像，首次提出旋转不变的自相似特征描述，在助力解决眼底图像配准具体问题的同时，也将为“模态无关的不变特征提取”这一多模态图像配准面临的共性科学问题开辟新思路。

内容获取失败，请点击重试