蜂窝划分的结合模型与拓扑约束的神经干细胞追踪问题的研究

项目名称： 蜂窝划分的结合模型与拓扑约束的神经干细胞追踪问题的研究

项目编号： No.60875020

项目类型： 面上项目

立项/批准年度： 2009

项目学科： 金属学与金属工艺

项目作者： 汤春明

作者单位： 哈尔滨工程大学

项目金额： 28万元

中文摘要： 为了解决海量神经干细胞序列图像分割与追踪中的一些挑战性问题，首先应用改进的SIFT实现了序列图像的自动配准，然后提出一种基于水平集和局部灰度阈值的分割算法，不仅能有效分割细胞与背景对比度低的图像，还能准确分离粘连、团簇细胞，并尽可能地保留细胞的形态特征和位置信息。该算法测试了两个图像序列，共170多帧。分割成功率最高可达72.86%，较传统的水线、Level Set提高了30%~40%。以拓扑约束法为基础，提出了一种蜂窝划分的多帧及原图反馈修正的细胞追踪方法，用扩展Kalman粒子滤波器完成了对活跃细胞的预测与追踪。该算法拓宽了拓扑约束的适用性，能有效追踪细胞分布疏密不均、团簇问题严重及细胞于帧间游动速度快的序列图像，并且大幅度地修正了分割错误，降低了追踪对分割的依赖性，最终使追踪准确率有很大地提高。该算法共测试了三个图像序列，共260多帧，共141个细胞。该方法的细胞追踪准确率最高可达95.83%，最低为80%，比拓扑约束法的追踪准确率分别提高了25.61%、77.14%和45.83%，达到了项目预期目标。

中文关键词： 序列图像；分割；拓扑约束；蜂窝划分；多帧反馈校正

英文摘要： To solve some challenging problems of segmentation and tracking large population of neuron stem cells in image sequences, we apply improved SIFT algorithm to accomplish images’automatic registration in sequences firstly. We have then proposed a segmentation algorithm based on level set combined with local gray threshold, which can not only effectively segment the images in low contrast, but accurately separate the adherent and clustered cells. It simultaneously keeps cells’shapes and locations as much as possible. Two image sequences have been tested, 170 frames totally. The maximal accuracy can approach to 72.86%. Compared with watershed and traditional level-set algorithms, the accuracy of the presented algorithm has been improved by 30%~40%. Based on topological constraints algorithm, we propose a novel cells’tracking method based on cellular partition combined with feedback correction of multi-frames and original images. And we devise a cell tracking feedback system, after applying Extended Kalman Particle Filter in prediction and tracking of the active cells. This system can expand the applicability of the topological constraints, and effectively track cells in the status of uneven density distribution, seriously clustering and moving fast. Because it may greatly correct segmentation error, it reduces the tracking dependence of segmentation. Three image sequences have been tested with 260 frames and 141 cells totally. The maximal tracking accuracy approach to 95.83% and the minimal one is no less than 80%. Compared with the topology constraints algorithm, the tracking accuracy has been improved 25.61%, 77.14% and 45.83% respectively, which has achieved the project’ anticipated target.

英文关键词： image sequences; segmentation; topology constraints; cellular partition; multi-frame feedback correction