基于图半监督学习的医学图像检索  

Medical Image Retrieval by Graph-based Semi-supervised Learning

作　　者：吴梦麟[1,2] 陈强[1] 孙权森[1] 

机构地区：[1]南京理工大学计算机科学与技术学院,南京210094 [2]南京工业大学电子与信息工程学院,南京211816

出　　处：《计算机辅助设计与图形学学报》2013年第9期1354-1360,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：自然科学基金(60805003)

摘　　要：针对医学图像检索中底层特征不能完整地描述图像的高层语义的问题,提出一种基于图的半监督学习框架的医学图像检索算法.首先根据图像之间的距离关系构建图模型,并在标记传播过程中加入密度相似性约束,得到查询图像的类别归属度,即图像的视觉语义表示;然后提取图像分块SIFT特征,用词袋进行描述,以获取图像的局部特征;最后设计了结合视觉语义和局部特征的相似性度量准则.在ImageCLEFmed上的实验结果表明,该算法能够有效地表达图像的视觉语义,检索效率优于单一底层特征检索.As low level features can not reflect the high level semantic in medical image search, an image retrieval algorithm is proposed by graph-based semi-supervised learning frame. Firstly, a graph model is constructed by distance between images, and density similarity constrained in the label propagation progress is added to get the membership degree of query images, called visual semantic representation; then the dense SIFT feature of the image blocks is extracted and described with bag of visual words, in order to get the local feature; Finally, a combination of visual concept and local feature strategy is designed for similarity measurement. Experimental results of ImageCLEFmed database demonstrate that the proposed algorithm represents the visual semantic of images effectively, and achieves a better retrieval performance than single low level feature.

关 键 词：基于内容的医学图像检索 基于图的半监督学习 视觉语义 词袋 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]