基于神经网络技术自动筛选Monastrol抑制剂的算法研究  

Classification approach based on neural network for high content analysis in Monastrol suppressor screens

作　　者：胡艺[1,2] 章东[1,2] 张喆[1,2] 葛云[2] 周晓波 

机构地区：[1]模式识别国家重点实验室,中国科学院自动化研究所,北京100080 [2]南京大学声学研究所,电子科学与工程系,南京210093 [3]Harvard Center for Neurodegeneration and Repair-Center for Bioinformatics,Harvard Medical School

出　　处：《南京大学学报（自然科学版）》2009年第1期39-47,共9页Journal of Nanjing University（Natural Science）

基　　金：教育部新世纪优秀人才计划项目(06-0450)

摘　　要：带自动荧光显微镜的HCS(high content screening)系统是一种新兴的显微摄影、筛选和处理系统.HCS系统在摄影过程中会产生大量数据,人工筛选和识别费时费力.本文为了进行Monastrol抑制剂的筛选,基于前馈式神经网络技术研究了一种对HCS系统摄影的大量细胞同时进行特征提取和细胞显型识别的自动算法.在得到各个通道分离的图像后,对不同通道图像进行并行预处理,并采用神经网络和逻辑运算相结合的算法进行处理.我们将该方法运用于Monastrol抑制剂的筛选中,并将结果与人工识别结果进行分析比较.相比较前人提出的multi-phenotypic mitotic analysis(MMA)算法自动识别的正确率得以提高,可更好评估抗癌药剂效用.HCS（high content screening） viaautomated fluorescent microscopy is a powerful technology for its effective expression of cellular processes. This technique obtains specific phenotypic information of cellular images by employing selected fluorescence probes. However, it generates great amount of image datasets at one time. which leads to difficulty in handling and analyzing. Manual recognition makes the processing extremely time- consuming and subjective. Consequently, there is an urgent need to develop effective automatic tool for processing and analyzing HCS data. Previously a multi-phenotypic mitotic analysis （MMA） algorithm was put forward to distinguish and label cells at different phases in cell division from three-channel acquisitions. Although MMA recognizes different mitotic phases well, the classification of bipolar and monoaster cells is not satisfying with an accuracy of about 85%. In addition, too much noise and rough extraction of cells make the classification of bipolar and monoaster cells not as well as expected. In order to better screen Monastrol suppressors, this article proposes an automatic classification approach for simultaneous feature extraction and cell phenotype recognition of monoaster and bipolar ceils in HCS system, based on a combination of neural network and logic operation. A parallel preprocessing and a hybird algorithm of BPNN and logic operations follow the acquisition of separated channel images. The purpose of this work is to improve the classification accuracy of bipolar and monoaster cells. Considering the sensitivity to noise of MMA algorithm, it is necessary to find a classification method with higher noise tolerance. We apply an algorithm based on a Back Propagation Neural Network （BPNN） in simultaneously extracting features and analyzing ceil phenotypes obtained by HCS system for improving the accuracy of bipolar and monoaster cell recognition, reducing the sensitivity to noise, as well as improving the evaluating drug efficiency. The validity of this approac

关 键 词：高容量分析 高容量筛选 图像分析 神经网络 模式识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]