同模态与多模态的眼底图像配准

发布时间：2019-04-11 18:48

【摘要】：眼底疾病种类繁多,常见的眼底疾病有视网膜血管病变,炎症性病变,黄斑性病变,视网膜病变等。眼底疾病十分影响人们的日常生活,而且白内障、青光眼、老年黄斑变性、糖尿病视网膜病变是四大致盲诱因,给人们日常生活带来了极大的不便。眼底图像是目前用来诊断眼底疾病的主要方式,无赤光眼底图像和荧光素眼底血管造影图像是临床诊断采用的最广泛的两类图像。由于一幅图像成像的视野有限,同模态眼底图像配准能合成更大的视野,确定病灶位置。多模态眼底图像的信息互补,多模态眼底图像配准能为医生提供更准确的病灶信息。然而目前这两类图像的配准是通过医生脑海的“人工剪影”实现的,这对医生的要求较高,且诊断时间长,漏诊和误诊的几率也比较大。为了研究眼底图像的配准,本论文研究了图像配准的基本理论,分析目前存在的图像配准算法的优缺点。将眼底图像的配准分为同模态的眼底图像配准和多模态的眼底图像配准。本论文将经典的图像配准算法SIFT用于同模态眼底图像的配准,调整SIFT算法中的相关参数,使SIFT算法能较好地适用于对比度较低的眼底图像。而多模态眼底图像间的灰度差异是非线性的,SIFT算法不再适用。本论文根据SIFT的每个步骤,分析SIFT算法不适用多模态眼底图像配准的具体原因,针对分析的原因,本论文提出了基于旋转不变距离的多模态眼底图像配准。首先用Harries提取眼底血管比较稳定的特征点,然后建立旋转不变描述子,接着计算特征点间的旋转不变距离,根据旋转不变距离匹配特征点,最后RANSAC消除误匹配,估计变换矩阵,实现图像融合。本论文研究的眼底图像数据来源于南ft医院,分别采集病理无赤光眼底图像和荧光素眼底血管造影图各80幅,健康的无赤光眼底图像和荧光素眼底血管造影图各80幅,总共320幅眼底图像。采用均方误差(RMSE)对同模态的眼底图像配准和多模态的眼底图像配准算法进行误差分析。根据实验结果分析,基于SIFT的同模态眼底图像配准的平均误差为2.40±0.5个像素,本论文提出的基于旋转不变距离的多模态眼底图像配准的平均误差为0.89±0.72个像素,匹配成功率为92.4%,并且算法性能不受图像旋转的影响。本文证明了经典的SIFT算法适用于同模态眼底图像配准,也证明了本论文提出的基于旋转不变距离的多模态图像配准算法克服了多模态眼底图像间的非线性灰度差异,实现了多模态眼底图像的配准。
[Abstract]:There are many kinds of fundus diseases, such as retinal angiopathy, inflammatory disease, macular disease, retinopathy and so on. Fundus diseases affect people's daily life very much, and cataract, glaucoma, senile macular degeneration and diabetic retinopathy are the four major causes of blindness, which bring great inconvenience to people's daily life. Fundus images are the main methods for the diagnosis of fundus diseases. Fundus-free fundus images and fundus fluorescein angiography images are the two most widely used images in clinical diagnosis. Because the field of view of one image is limited, the same mode fundus image registration can synthesize a larger field of view and determine the location of the lesion. The information of multimodal fundus images is complementary, and multi-modal fundus image registration can provide more accurate information for doctors. However, at present, the registration of these two kinds of images is realized by "artificial silhouette" in the mind of the doctor, which requires a higher doctor, and the diagnosis time is long, and the probability of missed diagnosis and misdiagnosis is also high. In order to study the fundus image registration, this paper studies the basic theory of image registration, and analyzes the advantages and disadvantages of the existing image registration algorithms. The fundus image registration is divided into the same mode fundus image registration and multi-modal fundus image registration. In this paper, the classical image registration algorithm SIFT is applied to the same mode fundus image registration, and the correlation parameters of the SIFT algorithm are adjusted to make the SIFT algorithm suitable for the fundus images with low contrast. However, the grayscale difference between multimodal fundus images is nonlinear, and the SIFT algorithm is no longer applicable. According to each step of SIFT, this paper analyzes the reasons why the SIFT algorithm is not suitable for multi-modal fundus image registration. Aiming at the analysis reasons, this paper proposes the multi-modal fundus image registration based on rotation invariant distance. Firstly, Harries is used to extract stable feature points of fundus vessels, then rotation invariant descriptor is established, then rotation invariant distance between feature points is calculated, feature points are matched according to rotation invariant distance, finally, RANSAC is used to eliminate mismatches and estimate transformation matrix. Realize image fusion. The fundus image data obtained in this paper were collected from South ft Hospital. The pathological fundus non-red fundus images and fundus fluorescein angiography (FFA) images were collected respectively, and the healthy non-red fundus images and fundus fluorescein angiography (FFA) images were each 80. A total of 320 fundus images. The mean square error (RMSE) was used to analyze the errors of the same mode fundus image registration algorithm and the multimodal fundus image registration algorithm. According to the experimental results, the average error of the same mode fundus image registration based on SIFT is 2.40 卤0.5pixels, and the average error of multimodal fundus image registration based on rotation invariant distance is 0.89 卤0.72 pixels. The matching success rate is 92.4%, and the performance of the algorithm is not affected by the image rotation. In this paper, the classical SIFT algorithm is proved to be suitable for the same-mode fundus image registration, and it is also proved that the multi-modal image registration algorithm based on rotation invariant distance can overcome the nonlinear gray-scale difference among the multi-modal fundus images. The registration of multimodal fundus images is realized.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R770.4;TP391.41

【相似文献】