Protein Tracking By CNN-Based Candidate Pruning And Two-Step Linking With Bayesian Network
Protein Tracking By CNN-Based Candidate Pruning And Two-Step Linking With Bayesian Network
复制标题
DOI：
10.1109/mlsp.2019.8918873
复制
发表时间：
2019-10
期刊：
影响因子：
--
通讯作者：
Mariia Dmitrieva;Helen L. Zenner;Jennifer H. Richens;D. Johnston;J. Rittscher
中科院分区：
文献类型：
--
作者：
Mariia Dmitrieva;Helen L. Zenner;Jennifer H. Richens;D. Johnston;J. Rittscher
蛋白质运输在理解许多生物过程和疾病中起着至关重要的作用。蛋白质囊泡的自动跟踪是具有挑战性的，因为它们的行为不稳定，外观变化和视觉混乱。在本文中，我们提出了一种新的跟踪方法，它利用了两步链接过程，利用概率图形模型来预测tracklet链接。囊泡最初是在候选选择过程的帮助下检测到的，其中候选者是通过多尺度斑点增强滤波器来识别的。随后，这些候选者被修剪并通过轻量级卷积神经网络进行选择。在连接阶段，基于距离和检测分配来形成轨迹线，该检测分配通过组合优化算法来实现。通过贝叶斯网络实现的概率模型用于推断应链接哪些轨迹。跟踪结果共聚焦荧光显微镜数据的蛋白质贩运上皮细胞。所提出的方法实现了一个均方根误差（RMSE）为1.39的囊泡定位和α为0.7代表的程度的轨道匹配与地面真理。所提出的方法也评估对国家的最先进的“Trackmate”的框架。
Protein trafficking plays a vital role in understanding many biological processes and disease. Automated tracking of protein vesicles is challenging due to their erratic behaviour, changing appearance, and visual clutter. In this paper we present a novel tracking approach which utilizes a two-step linking process exploiting a probabilistic graphical model to predict tracklet linkage. The vesicles are initially detected with help of a candidate selection process, where the candidates are identified by a multi-scale spot enhancing filter. Subsequently, these candidates are pruned and selected by a light weight convolutional neural network. At the linking stage, the tracklets are formed based on the distance and the detection assignment which is implemented via combinatorial optimization algorithm. A probabilistic model, realised through a Bayesian network, is used to infer which tracklets should be linked. Tracking results are presented for confocal fluorescence microscopy data of protein trafficking in epithelial cells. The proposed method achieves a root mean square error (RMSE) of 1.39 for the vesicle localisation and α of 0.7 representing the degree of track matching with ground truth. The presented method is also evaluated against the state-of-the-art “Trackmate“ framework.