基于双字典学习的眼底图像血管分割

Fundus image blood vessel segmentation via joint dictionary learning

为辅助诊断眼底疾病和部分心血管疾病,本文提出一种基于双字典学习和多尺度线状结构检测的眼底图像血管分割方法。首先在HSV颜色空间利用伽马矫正均衡眼底图像的亮度,并在Lab颜色空间采用CLAHE算法提升图像对比度,再采用多尺度线状结构检测算法突出血管结构得到增强后的特征图像;然后利用K-SVD算法训练特征图像块和对应的手绘血管标签图像块,得到表示字典和分割字典,采用表示字典得到新输入特征图像块的重构稀疏系数,由该系数和分割字典获得血管图像块;最后进行图像块拼接、噪声去除和空洞填充等后处理得到最终分割结果。在DRIVE和HRF数据库测试,利用准确率、特异度、敏感度等八种评估指标来检验分割性能。其中,平均准确率分别达0.958 2和0.951 5,平均特异度分别达到0.982 6和0.967 1,平均敏感度分别达到0.709 5和0.762 6,表明该方法具有较好的分割性能和通用性。

In order to assist the diagnoses of fundus diseases and some cardiovascular diseases,this paper proposes a fundus image blood vessel segmentation method via joint dictionary learning and multi-scale line structure detection.Firstly,brightness is adjusted and balanced by gamma correction in HSV color space,contrast is improved via CLAHE algorithm in Lab color space,and multi-scale line structure detection algorithm is used to enhance the blood vessel structures and get the feature maps.Then,the representation dictionary and segmentation dictionary are trained simultaneously by K-SVD algorithm from the feature blocks and its corresponding manually annotated vessel label blocks.The reconstructed sparse coefficients of newly input enhanced feature blocks are obtained with the representation dictionary,and the blood vessel blocks are segmented by these coefficients and segmentation dictionary.Finally,the blood vessel result is obtained via image blocks stitching,noise removal and hole filling algorithms.Our method is tested on DRIVE and HRF databases to evaluate the segmentation performance in accuracy,sensitivity,specificity and other five metrics.The average accuracy rate reaches 0.958 2 and 0.951 5 respectively,the average specificity reaches 0.982 6 and 0.967 1 respectively,the average sensitivity reaches 0.709 5 and 0.762 6 respectively,which indicates that our method has good segmentation performance and versatility.