聚类核值相似区特征点的医学影像分类  

Medical image classification based on cluster univalue segment assimilating nucleus feature points

作　　者：李博[1,2] 曹鹏[1,2] 栗伟[1,2] 赵大哲[1,2] 

机构地区：[1]东北大学信息科学与工程学院,沈阳110004 [2]东北大学医学影像计算教育部重点实验室,沈阳110004

出　　处：《中国图象图形学报》2013年第10期1322-1328,共7页Journal of Image and Graphics

基　　金：国家自然科学基金项目(61001047);中央高校基本科研业务费专项基金项目(N110618001)

摘　　要：针对传统方法通常选取角点或极值点作为特征点,忽略了局部纹理变化从而影响医学影像分类性能的问题,提出一种新的特征点检测和描述方法,并应用Bag-of-Keypoints模型实现医学影像分类。首先改进自适应的Kmeans对影像进行像素级聚类,构建核值相似区并选取邻域内聚类分布变化急剧的像素点作为特征点;然后在极坐标系中定义特征点描述符并生成视觉词典,通过视觉词直方图描述影像;最后利用直方图交集方法度量影像间的相似度来扩展KNN(K-nearest neighbor)完成分类。遵循IRMA(image retrival in medical appication)的医学影像类别编码标准,严格选择实验数据,结果表明该算法较传统方法 F1值平均提高4.5%,对于不同类别影像效果更加稳定鲁棒,从而更好地满足临床应用需求。Traditional methods usually use corners and extreme points as feature points, and ignore the changes of texture so that the performances of the medical image classification are affected. A new feature point detection and description method is provided for medical image classification task using the Bag-of-Keypoints model. First, adaptive K-means is used to cluster images on the pixel-level, and the points where the clustering distribution in its univalne segment assimila- ting nucleus （USAN） changes rapidly are selected to be the features points. Second, the descriptor is defined in a polar coordinate system and a virtual dictionary is constructed in order to describe the image by virtual word histogram. Last, histogram intersection is used to measure the similarity between images, and K-nearest neighbor （KNN） is extended by it to finish the classification. Image retrieval in medical applications （IRMA） medical image classification code is strictly followed when experimental data is selected. The results show that the average Ft value increased 4. 5% than traditional classification ; our method is more stable and robust for different classes of image, and meets the needs of clinical applica- tion better.

关 键 词：影像分类 特征点 Bag—of-Keypoints模型 自适应聚类 核值相似区 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]