【作者】 王玉亮；

【导师】 廖文和；

【作者基本信息】 南京航空航天大学 ， 机械工程， 2012， 博士

【摘要】 医学影像学及相关技术已广泛用于医学领域，并发展为医学临床诊疗中的客观依据，医学影像的处理与分析技术作为辅助诊疗的关键，具有重要的临床和研究价值。眼底图像是眼科中通过眼底照相机获取的一种标准的客观诊断影像，其中，眼底是位于内眼后部组织结构（视网膜、脉络膜、视神经和黄斑等）的统称。眼底疾病（糖尿病视网膜病变、老年性黄斑变性和青光眼等）可导致视力下降，是致盲的首要原因，为此，眼底图像的处理与分析对糖尿病、高血压等眼底病变以及黄斑病变、眼底动脉硬化和视网膜病变等各种眼底疾病的早期发现、诊断、治疗以及辅助诊疗有着重要的意义。论文按照眼底图像获取、眼底图像预处理、眼底图像配准融合拼接、眼底重要目标分割与测量、眼底图像处理与分析系统实现这条主线展开研究。主要研究内容和创新成果如下：（1）针对眼底照相机对人眼眼底的成像原理，由Gullstrand I号模型眼参数建立了人眼光学模型，由眼底照相机成像光路设计建立了眼底照相机光学模型，分析了人眼和眼底照相机成像光路中存在的像差和畸变，提出了基于眼底照相机标定的眼底图像几何畸变校正算法及基于RGB与HSI彩色图像空间变换和频域同态滤波的彩色眼底图像辐射量畸变校正算法，可有效校正眼底图像中最大为20像素左右的几何畸变，有效校正彩色眼底图像的辐射量畸变，保留彩色眼底图像的细节信息，提高后续参数测量的精度；针对人眼光学模型、眼底照相机成像光路的像差和畸变特点，引入了一种基于盲解卷积和Lucy Richardson的彩色眼底图像复原方法，可有效改善彩色及灰度眼底图像的清晰度，分别提高16.4%和15.8%，保证了后续处理与分析的精度；并针对不同的图像处理与分析工作进行针对性的眼底图像预处理。（2）针对眼底照相机成像的视场局限以及眼底图像对比度低、光照不均匀、不同视场间存在几何畸变等缺点，应用了一种基于SSDA的眼底图像半自动配准方法，可交互式半自动配准低对比度的眼底图像；提出了一种基于改进Harris角点检测的眼底图像自动配准算法，可配准对比度较好的眼底图像，平均配准误差约为2.5像素；提出了一种基于SIFT特征的眼底图像自动配准算法，改进了SIFT算法中的特征检测及匹配参数，改进了特征点匹配后的点对提纯算法，优化了图像配准模型，实现了多幅不同视场正常或病变眼底图像的自动快速、高精度鲁棒配准，平均配准误差不超过2.1像素，但耗时较多，适用于硬件加速；提出了一种基于SURF特征的眼底图像自动配准算法，改进了SURF算法中的参数及相关阈值，改进了特征点匹配后的点对提纯算法，优化了配准模型，实现了基于SURF-128的多幅不同视场正常或病变眼底图像的自动快速、高精度鲁棒配准，平均配准误差不超过2.2像素，适用于常规计算机。（3）分析了不同视场眼底图像间的空间关系及配准策略，推导了二次多项式变换更适合于眼底视网膜近似二次曲面的全景图像配准，应用了基于像素（线性、非线性）、基于拉普拉斯等金字塔、基于小波的RGB空间和Lab空间的彩色眼底图像融合方法以保证眼底全景图像融合质量，验证了眼底图像主客观评价的有效性，改进了眼底图像的清晰度评价函数，结合质量评价指标对比分析了各种彩色眼底图像融合方法，确定了各种图像融合方法的参数；提出了一种基于先验知识的眼底全景图像自动拼接算法，实现了基于各眼底视场中视盘、血管等先验知识以及加入中间拼接图像步骤的快速、高精度（像素级）的眼底全景图像自动配准与拼接。（4）针对眼底图像中重要目标（血管、视盘和视杯等）的特点，提出了一种基于高斯核函数与Hessian矩阵的眼底血管多尺度分割算法和基于大尺度血管的视盘自动定位与Hough变换的圆拟合视盘分割相结合的参数测量算法，可快速、有效检测得到眼底图像中亚像素精度的血管中心线坐标、中心线方向、边界点坐标、宽度和长度等信息，并可进行快速高精度的视盘分割和参数测量；提出了一种基于活动轮廓的视盘视杯分割方法以有效进行视盘视杯分割和参数测量，结合眼底图像畸变校正，可提高分割与测量的精度。（5）针对眼底图像处理与分析系统的需求，深入研究了基于CUDA（硬、软件系统）的眼底图像快速自动配准与拼接中的关键技术，提出了一种基于CUDA的眼底图像快速自动配准与拼接算法，用CUDA加速了眼底图像的同态滤波增强、改进算法参数的SIFT（或SURF）特征点检测、匹配和匹配提纯算法、最优模型变换矩阵计算、配准与拼接，实现了多幅眼底图像的快速、高精度自动配准与拼接，算法运行速度与仅用CPU运行相比提升了10~30倍，保证精度的前提下算法效率得到极大提高；建立了一套可达到临床应用要求的眼底图像处理与分析系统，实现了眼底图像获取、配准与融合拼接以及视盘和血管等重要目标定位、分割与参数测量，为眼底疾病预防和辅助诊疗提供了有力工具。更多还原

【Abstract】 Medical imaging and correlation technique widely used in medical field has been developed as theobjective basis of clinical diagnosis and treatment, medical image processing and analysis technologyas the key auxiliary diagnosis and treatment, which has important clinical and research value. Fundusimage obtained by fundus camera is a standard objective diagnostic imaging in ophthalmology, andthe fundus is located in internal ocular. Fundus oculi disease is the leading cause of blindness, whichcan cause vision loss. Therefore, fundus image processing and analysis is of great significance to earlydetection, diagnosis, treatment, auxiliary diagnosis and treatment of various kinds of retinal disease,such as diabetes, hypertension, maculopathy, arteriosclerosis, retinopathy, and so on.The paper focuson fundus image acquisition, image preprocessing, image registration, image fusion and mosaic,important target segmentation and measurement of fundus image, and the realization of fundus imageprocessing and analysis system. The main points are as follow:(1) According to the fundus camera imaging principle, establish the human eye optical model byGullstrand I model parameters, and establish fundus camera optical model by optical design, thenanalyses the existing aberration and distortion of human eye and fundus camera imaging opticalsystem, fundus image geometric distortion correction algorithm based on fundus camera calibrationand color fundus image radiation distortion correction algorithm based on RGB and HSI color spacesand homomorphic filter are proposed, which can effectively correct the geometric distortion (about20pixels) and radiation distortion of color fundus image, preserve color fundus image detailsinformation, and improve the precision of parameters measurement; According to the characteristicsof human eye optical model and fundus camera imaging optical system, proposes a color fundusimage restoration method based on blind deconvolution and Lucy Richardson, which can effectivelyimprove the color or gray fundus image resolution, increased by16.4%and15.8%respectively,improve the accuracy of subsequent processing and analysis. And the specific fundus imagepreprocessing methods are used to dealing with specific problems.(2) In the basis of the fundus camera imaging field limited, low contrast, illumination uneven andgeometric distortion, a semi-automatic image registration method based on SSDA is applied toregister low contrast fundus images; And propose an automatic fundus image registration algorithmbased on improved Harris corner detection to register better contrast fundus images; An automaticfundus image registration algorithm based on SIFT features is proposed, which improves thealgorithm parameters of feature detection, matching and purification, optimize the image registration model, can automatic register fundus images quickly with high precision, and the average registrationerror is no more than2.1pixels; An automatic fundus image registration algorithm based onSURF-128features is proposed, which improves the algorithm parameters of feature detection,matching and purification, optimize the image registration model, can automatic register fundusimages quickly with high precision, and the average registration error is no more than2.2pixels.(3) The spatial relationship and registration strategy in different fundus images fields is analysed.And it is derived and verified that the quadratic polynomial transformation is suitable for the funduspanoramic image registration. And color fundus image fusion method based on pixel nonlinear,Laplacian pyramid and wavelet (in RGB and Lab space) is applied to ensure the fundus panoramicimage quality. The effectiveness of the fundus image subjective and objective evaluation is verified,and improve the fundus image clarity evaluation function, determine the various parameters of theimage fusion method by comparison and analysis of the various color fundus image fusion method.An automatic fundus image panoramic mosaic algorithm based on the prior distribution knowledge ofoptic disk and blood vessels is proposed, and adding intermediate mosaic image steps, can automaticachieved high precision fundus panoramic image registration quickly.(4) In the basis of the characteristics of fundus image, analyzes the important target for retinalimage (blood vessels, optic disk and optic cup, et al.), a retinal vascular multi-scale segmentationalgorithm based on gaussian kernel function and Hessian matrix is proposed to quickly and effectivelydetect retinal vascular center line coordinates, line direction, boundary point coordinates, width andlength, and other information in fundus image. An automatic optic disk location algorithm based onlarge scale vessel segemention and optic disk segemention method based on edge (or area)segmentation and Hough transform circle fitting is proposed to segment and measure optic disk withhigh precision. Optic disk and optic cup segmentation method based on active contour is proposed,which can accuracy segment and measure optic disk and optic cup with image distortion correction.(5) To meet the requirements of fundus image processing and analysis system, a fast automaticfundus image registration and mosaic algorithm based on CUDA is proposed, and the algorithm speedupgrade10to30times. To meet the needs of clinical application, the fundus image processing andanalysis system is established, and realized fundus image acquisition, registration and fusion, mosaic,segmentation and parameter measurement of optic disk and vessels, and provide a powerful tool forretinal disease prevention, auxiliary diagnosis and treatment.更多还原