【作者】 潘晨； 闫相国； 郑崇勋；

【Author】 Pan Chen; Yan xiangguo; Zheng Chongxun(Key Laboratory of Biomedical Information Engineering of MOE,Xi’an Jiaotong University,Xi’an, 710049,China)

【机构】 西安交通大学生物医学信息工程教育部重点实验室；

【摘要】 提出一种新的血细胞图像分割算法,结合无监督和有监督模式识别技术,利用颜色快速提取感兴趣的有核细胞。首先通过mean-shift过程寻找RGB颜色空间中的细胞核、红细胞和背景聚类峰(局部密度最大区域),其中的细胞核图像区域经过适当形态学膨胀后可以得到部分胞浆像素;然后细胞核聚类峰和部分胞浆颜色组成正特征子集,而红细胞和背景聚类峰附近颜色组成负特征子集,训练一个两分类SVM,得到的分类模型随后对图像的颜色空间向量分类,实现细胞区域整体提取。通过颜色量化手段,能够显著减少训练集的颜色向量数量,实现SVM实时训练和分类。借助于mean-shift鲁棒的特征空间分析性能和SVM出色的小样本学习推广能力,该方法对图像颜色变化、染色条件差异等鲁棒性强,无过度分割现象,分割速度和效果均优于流域变换方法。骨髓和外周血涂片的分割试验证明了方法的有效性,已应用于实际图像处理系统。更多还原

【Abstract】 A blood cell image segmentation method is proposed which combines mean-shift algorithm and support vector machine (SVM) to extract nucleated cells. Firstly, three mean-shift procedures were used to find clustering mode (local maxima of frequency)of nucleus, mature erythrocytes and background respectively. Then the region of nucleus would be inflated in image in order to get a part of color information of cytoplasm around nucleus. A part of color of cytoplasm and nucleus would comprise the positive training subset, and color of mature erythrocytes and background do the negative training subset. A two-class SVM could be trained with the training set and generate a classification model. This model is subsequently used to classify image pixels. By the means of color quantization, the number of color vectors in training set would be reduced significantly, so the training speed is close to real time. Due to mean-shift procedure has robust performance to feature space analysis and SVM has good generalization performance trained with small samples, the new approach is robust to varied preparation and illumination. The approach is better in segmentation effect and shorter in consuming time, compared with classical watershed algorithm. Experimental results demonstrate the efficiency of the new method.更多还原