改进的快速FCM及SVM实现糖网白色病灶的自动检测

摘要

To develop an automated diabetic retinopathy (DR) screening system,an automatically detecting approach based on improved and fast FCM (IFFCM) as well as SVM was established and studied.First,color fundus images were segmented by IFFCM,and candidate regions of bright lesions were obtained.The median filter was added to objective function of FCM and the result of K-means clustering was used to initialize clustering centers of FCM,so the new algorithm overcome the shortcomings of high complexity and sensitivity to noise.Second,a two-level SVM classification structure was applied to classify the candidate regions.The bright lesions were picked up by features of candidate regions in stage one.Another group of features were used to discriminate hard exudates from cotton wool spots in stage two; as a result the automated detection of bright lesions in fundus images was accomplished.The approach was tested on a new set of 65 fundus images.With an image-based criterion,sensitivity of 100％,specificity of 95.0％ and accuracy of 98.46％ are achieved.Average sensitivity of 96.42％/97.15％ and average positive predict value of 90.03％/91.18％ are also achieved with a lesion-based criterion (hard exudates/cotton wool spots).Furthermore,the average time cost in processing an image is 35.56 seconds.Results suggest that the combination of the good result of coarse segmentation provided by IFFCM and higher recognition rate of SVM makes the results of automated detection better.It means that the proposed approach can efficiently detect bright lesions of DR from fundus images.%为构建基于眼底图像的糖尿病视网膜病变(糖网)自动筛查系统,提出一种基于改进的快速FCM(IFFCM)及SVM的糖网白色病灶自动检测算法.首先,利用改进的快速FCM算法,对彩色眼底图像进行粗分割获取糖网白色病灶候选区域,由于该算法将中值滤波添加到FCM算法的准则函数中,同时利用K-means算法的聚类结果对FCM进行聚类中心初始化,使得该算法克服了传统FCM算法计算复杂度高以及对噪声敏感的缺点；其次,采用两层级联分类结构的SVM对候选区域进行分类,即先利用SVM根据候选区域的特征集将白色病灶提取出来,再利用SVM根据另外的特征集将白色病灶中的硬性渗出与棉绒斑区分开,从而实现眼底图像中糖网白色病灶的自动检测.利用该方法对65幅眼底图像进行糖网白色病灶的自动检测,得到图像水平灵敏度100％,特异性95.0％,准确率98.46％；病灶区域水平(硬性渗出/棉绒斑)灵敏度96.42％/97.15％,阳性预测值90.03％/91.18％；平均一幅图像处理时间35.56 s.结果表明:将改进的快速FCM算法所提供的良好粗分割结果与识别率较高的分类器SVM相结合,使得对糖网白色病灶的自动检测结果较优,即该算法能够高效地自动检测出眼底图像中的糖网白色病灶.